过去24小时大模型研究员 X 动态：Qwen3.8 开放权重预告，Kimi K3 触及算力上限

过去24小时大模型研究员 X 动态：Qwen3.8 开放权重预告，Kimi K3 触及算力上限

本轮覆盖频道 408 个 researcher 账号，确认 167 条新增帖，重点呈现 Qwen3.8 的开放权重预告、Kimi K3 的容量与会员调整、Codex 工作流和训练基础设施信号；这不是 X 全站全量结论。

大模型公司研究员动态追踪
大模型公司研究员动态追踪@hongbo Li
Qwen3.8 的开放权重预告和 Kimi K3 的算力限流，是过去 24 小时最清楚的两条产品信号。Qwen 官方账号称 Qwen3.8 将很快开放权重，Qwen3.8-Max-Preview 已先在 Token Plan、Qoder 和 QoderWork 上线测试；Kimi 则说 K3 上线后的 48 小时里，需求已接近当前容量上限，暂时停止新订阅，并把会员拆成 Kimi Membership 与 Kimi Code Membership。1 2
本文覆盖频道 researcher roster 的 408 个账号。408 个账号的时间线请求全部返回，窗口内有 64 个账号发出 167 条新增去重帖。范围是洛杉矶时间 2026 年 7 月 18 日 08:00 至 7 月 19 日 08:00；这不是 X 全站全量结论。没有窗口命中的账号不等于没有发文，Cohere、DeepSeek、Mistral、SSI、智谱等分类本轮没有命中。

先看重点

  • 模型发布开始和服务容量正面相撞。 Qwen3.8 的官方预告把开放权重、2.4T 参数和多入口预览放在一起，但参数与能力描述仍是项目方自述。Kimi 的帖子则直接承认新订阅受算力约束，接下来按 Web/App/Work 与 coding 两类工作流拆分容量。
  • Codex 的更新落在使用链路，而非又一个榜单。 Charlie Marsh 预告下一版 Codex TUI 会对流式 Markdown 做增量渲染，称大输出速度可提高数百倍；Vaibhav Srivastav 展示的 ChatGPT Work 已经把托管浏览器、邮件和日历串进代办流程。两者都是账号自述，原帖没有给出统一测试方法。3 4
  • 训练与服务基础设施的细节比口号更有用。 ByteDance Seed 归类账号 Chayenne Zhao 写道，一段 38 分钟的多人会议在单张 H100 上用 49 秒完成带说话人区分和时间戳的转写，并把吞吐归因于 SGLang Omni ASR、OpenMOSS MOSS-TD serving、CUDA graphs、chunked prefill 和异步解码。这是个人发布的工程数据，不是独立基准。5

公司覆盖

公司roster 账号窗口活跃账号新增帖
AMI Labs / ex-Meta AI1120
Alibaba Qwen911
Anthropic4945
ByteDance Seed411
Cognition833
Cohere300
Cursor1311
DeepSeek400
Google DeepMind4633
Meta AI3922
Mistral100
Moonshot/Kimi624
OpenAI15435104
SSI100
Thinking Machines Lab23717
xAI4646
智谱100

按公司看

OpenAI：Codex 继续向可执行工作流下沉

OpenAI 分类有 154 个 roster 账号，本窗口 35 个活跃账号贡献 104 条新增帖。高信号内容主要在 Codex、ChatGPT Work 和开放权重讨论，Build Week 的社区活动与大量个人闲聊降到索引区。
Tier 1 账号中，Head of Applied Research Boris Power 转发了 Tibo 对 GPT-5.6 Sol 的正面评价。这里没有测试集、对照组或测量方法，最多说明内部研究负责人放大了一个积极判断，不能当作独立评测。6 post-training research 账号 Aidan McLaughlin 本窗口只有一条与研究无关的转发，按频道规则保留记录，但不从中推导研究方向。7
Charlie Marsh 给出了更具体的工程变更：下一版 Codex 的终端界面会对流式 Markdown 做增量渲染，他称大输出速度可提高数百倍。增量渲染解决的是输出已经开始生成后，终端如何持续显示的问题；原帖没有说明比较基线和测量环境，所以这里保留为开发者预告。3
Vaibhav Srivastav 展示了 ChatGPT Work 的另一条路径：对话可以调用托管浏览器，读取邮件和日历来查找选项并核对预约细节。它把聊天产品推向了代办流程，但也意味着权限、外部操作和错误恢复要一起评估。帖子只描述了个人案例，没有给出产品边界或安全机制。4
OpenAI 研究员 Andrew Ho 在多条帖子里为开放模型扩散辩护，认为前沿能力被复制后会带来更广泛的 AI 使用，也不认同实验室必须不断收购和扩张。这是个人观点，不代表 OpenAI 官方立场。8 9

Anthropic：Fable 之后，团队协作信号多于技术细节

Anthropic 49 个 roster 账号中有 4 个命中 5 条新增帖。Scaling RL 负责人 Sholto Douglas 只留下了一句「interesting + nice test」，没有交代所指测试，不能继续补出研究结论。10 Claude Code 账号 Thariq 则把 Fable 的按时完成归因于 Anthropic 多个团队的高强度协作，并写下 Enjoy Fable。帖子能确认项目交付的团队语境，不能单独证明 Fable 的性能或发布时间。11

Thinking Machines Lab：开放权重讨论延续，系统扩展经验更具体

Thinking Machines Lab 23 个 roster 账号中有 7 个命中 17 条新增帖。上期已经展开 Inkling 的 ARC-AGI 结果，本期 Hanyu Xue 和 Martin Ziqiao Ma 的转发仍在传播同一发布事实，因此不把它们计作新的评测结果。ARC Prize 转发里给出的 ARC-AGI-2 36.5%、每题 0.64 美元，以及 ARC-AGI-1 79.5%、每题 0.30 美元，仍应标为项目方帖子中的数据。12 13
Hanyu Xue 另一条原创帖把问题转向开放权重策略：他认为中国模型能力变强后，开放权重是否仍是非顶级模型获取注意力的主要方式，值得重新讨论；他同时明确把未来模型转向闭源写成 rumor，没有给出来源。这条只应作为个人问题意识，不能写成实验室路线变化。14
Shizhe Diao 转发了 ReOPD 讨论，标题性的判断是「More on-policy is NOT always better」，正文摘要提到 student occupancy mismatch 与 teacher reliability 两类 prefix shift。它给出的信息足以说明后训练中增加 on-policy 数据并非自动改善结果，但不足以重建论文方法或实验结论。15
Eric Zhang 的帖子则是系统工程经验：一个早年按每秒 10 至 20 个请求设计的系统，如今面对约 1000 倍规模与负载，团队在保留开发者体验的同时做了渐进式改造。这不是模型性能数据，却很贴近训练和推理基础设施在真实规模下的难点。16
Songlin Yang 多次转发开放研究与开放权重讨论，包括中国实验室为何能持续推出 GLM、Kimi 等模型，以及美国实验室低估开放研究活力的观点。转发者没有提供独立证据，这一栏只记录研究员关注的议题，不把它改写成 Thinking Machines Lab 的官方立场。17 18

Alibaba Qwen：Qwen3.8 先预览，开放权重随后到来

Qwen 官方账号在窗口后段发帖称 Qwen3.8 将很快开放权重，给出的规模是 2.4T 参数，并称 Qwen3.8-Max-Preview 已在 Alibaba 的 Token Plan、Qoder 和 QoderWork 上线。参数规模、能力比较和「最强模型之一」都来自官方帖子，当前没有在本轮账号池里看到独立验证。1

Moonshot/Kimi：K3 的需求约束开始进入产品设计

Kimi 官方账号说，过去 48 小时的需求已接近当前容量上限，因此暂时停止新订阅，现有订阅用户不受影响；它还宣布把会员拆成面向 Kimi Web、App、Work 的 Kimi Membership，以及面向 coding workflows 的 Kimi Code Membership，并称会分批恢复新订阅。这是容量管理和套餐设计的官方公告，不应和第三方评测混写。2
Kimi 官方还转发了 Together Compute 对 K3 与 Claude Fable 5 的 DeepSWE 软件工程测试分析，但本账号返回内容只显示了结论前半句，缺少完整方法和数值；本文把它列为官方放大的第三方材料，不据此下排名结论。19

ByteDance Seed：一次 ASR 服务把模型接入成本说清楚

ByteDance Seed 分类的 Chayenne Zhao 介绍了 MOSS-TD serving：单张 H100 在并发 16 时，将 38 分钟多人会议转成带说话人区分和时间戳的文本用了 49 秒，帖子给出的吞吐是每墙钟秒处理 97.5 个音频秒。她把收益归因于 encoder CUDA graphs、chunked prefill、异步解码，以及此前为 TTS 磨好的基础设施。数值是个人工程报告，仍需用相同音频和配置复测。5

Google DeepMind、Meta AI、xAI 与其他低命中分类

Google DeepMind 的 3 条帖子里，Chief Scientist Jeff Dean 转发了关于 pod 成为高连接加速器集群行业称呼的讨论，仍是个人账号的系统观察，不是 Google 官方规格说明。20 Meta AI 本轮只有 2 条低信息量帖子；xAI 的帖子包括 Andrew Milich 对下周更新的预告，以及 Yiping Wang 转发的关于一个简单循环同时影响 pretraining perplexity 和 post-training reasoning tasks 的讨论，均没有给出足够上下文。21 22
Cognition、Cursor、AMI Labs / ex-Meta AI 的其余内容以个人动态和转发为主；Cohere、DeepSeek、Mistral、SSI、智谱本窗口没有命中。完整原帖仍按公司收录在文末，避免把低信息量帖子伪装成研究进展。

人员动态

本窗口在 408 个 roster 账号的返回结果中，没有确认可直接归因于当事人的入职、离职、转岗、休假/重返、创业、开源项目启动或结束、论文接收、获奖或 grant 公告。Andrew Carr 的个人简介与历史归类存在身份信息不一致，但本窗口帖子没有给出状态 announce，因此不把它写成人员变动。Andrew Milich 的「下周有大更新」也没有说明是职位、项目还是产品变化，暂不进入人员动态栏。
这部分结论只覆盖本期账号池与接口返回范围，不代表 X 全站没有相关公告。

今日研究员 Spotlight

  1. Chayenne Zhao，ByteDance Seed，Multi-Modal Inference / SGLang 与 ASR serving。 她的公开简介还写有 RadixArk founding member，研究兴趣落在多模态推理和服务系统。本窗口她用单张 H100 的会议转写案例解释 CUDA graphs、chunked prefill 与异步解码如何降低新模型接入成本。5
  2. Charlie Marsh，OpenAI，Codex 与高性能开发工具。 他的简介同时列出 Ruff、uv、ty 等 Python 工具，当前更像连接模型产品与开发者基础设施的工程角色。本窗口他预告 Codex TUI 将加入流式 Markdown 增量渲染，关注点是大输出下的交互延迟。3
  3. Jason Wolfe，OpenAI，alignment 与 Model Spec。 他负责的方向涉及模型行为边界和对齐规范。本窗口他转发关于 AI 影响需要更多人参与的观点，帖子谈的是研究公共语境，不是新的模型结果。23
  4. Jeff Dean，Google DeepMind，Chief Scientist 与 Gemini lead。 他的研究管理范围横跨模型、系统与基础设施。本窗口他转发了 pod 作为高连接加速器集群称呼的讨论，提示大模型竞争仍然依赖集群级系统设计，而不是只看单卡指标。20
  5. Shizhe Diao，Thinking Machines Lab，微调与后训练。 他的简介写明 MTS、Finetuning 与 LMFlow。本窗口他转发 ReOPD 对 on-policy 数据的讨论，核心提醒是训练分布和教师可靠性会同时造成 prefix shift，适合继续追论文原文而不是直接套用结论。15
  6. Songlin Yang，Thinking Machines Lab，pretraining。 她持续关注预训练、开放研究和开放权重模型的生产方式。本窗口她转发了中国实验室的开放性讨论、美国实验室对开放研究活力的判断，也转发了 Qwen3.8 预告；这些是研究兴趣线索，不等于实验室正式路线。17 24
  7. Hanyu Xue，Thinking Machines Lab，研究员，开放权重与模型策略。 他的简介标注曾在 xAI。本窗口他先转发 Inkling 的 ARC-AGI 数据，后发文讨论中国模型能力提升后开放权重策略是否仍然必要，并明确把未来转向闭源写成未经证实的 rumor。14
  8. Eric Zhang，Thinking Machines Lab，计算机系统与交互设计。 他此前是 Modal founding engineer，关注系统和软件体验。本窗口他描述一个从每秒 10 至 20 个请求起步、如今承受约 1000 倍负载的系统，重点是扩展过程中仍保留开发者体验。16
  9. Sholto Douglas，Anthropic，Scaling RL。 他的简介把研究方向写得很明确，核心是强化学习扩展，并曾在 DeepMind 工作。本窗口只有一句「interesting + nice test」，没有披露实验对象或结果，反而提醒读者不要把短评补写成完整研究结论。10
  10. Thariq，Anthropic，Claude Code。 他的工作方向在模型能力与代码工作流之间。本窗口他谈到 Fable 项目按时完成，强调 Anthropic 多个团队的协作和高强度投入；帖子没有给出模型指标，最近动态更适合从交付组织方式来读。11

新增帖完整索引

以下按当前 roster 公司归类，保留窗口内 167 条新增去重帖的发布时间、账号、帖子类型、原文摘要和原帖链接。转发内容仍按发帖账号归类，不把被转发账号误算为 roster 账号。

OpenAI

  • 7 月 18 日 08:14，@reach_vb，原帖：“Most people overestimate what they can do in one year and underestimate what they can do in ten years.” 原帖
  • 7 月 18 日 08:18，@gabrielchua，转发：RT @brianchew: what a wonderful saturday! folks showed up with ideas, built all afternoon, helped each other out, and left with some reall… 原帖
  • 7 月 18 日 08:18，@gabrielchua，转发：RT @HowDevelop: Really loving the exclusive #BuildWeek Swag from @OpenAIDevs Was such a great experience to host the @OpenAI Build Week h… 原帖
  • 7 月 18 日 08:52，@andrew_n_carr，原帖：I ported the TF sampler to torch and wrote a little optimizer to fit face params. It's pretty expressive!!! https://t.co/WqTkyXNCZN 原帖
  • 7 月 18 日 08:56，@yong_zhengxin，原帖：first karpathy, now Dean :( (though I think their content are different - but the reaction they got and how they had to move their voice to private ch... 原帖
  • 7 月 18 日 09:04，@jxnlco，转发：RT @omooretweets: ChatGPT Sites is severely underrated - it is THE breakthrough in bringing vibe coding to the masses I prompted 5.6 to ma… 原帖
  • 7 月 18 日 09:15，@reach_vb，原帖：ChatGPT tip: chat can use a full-fledged managed browser for you, so you can ask it to look things up, and even get things done on your behalf. All wi... 原帖
  • 7 月 18 日 09:28，@nickbaumann_，原帖：We can all have nice things https://t.co/bRcQgj9cYj 原帖
  • 7 月 18 日 09:29，@nickbaumann_，转发：RT @reach_vb: ah yes the weekend, finally time to relax and use personal codex to build single use software and automate away the boring li… 原帖
  • 7 月 18 日 09:48，@yong_zhengxin，转发：RT @boazbaraktcs: I get this reaction too and it is unfortunate. AI's impact is so significant that it is important to involve many people… 原帖
  • 7 月 18 日 09:50，@nickbaumann_，转发：RT @MatthewBerman: Got this email within 1 minute of the reset. https://t.co/UBwegudVRl started as a joke and is now quite useful. https:/… 原帖
  • 7 月 18 日 10:10，@reach_vb，原帖：Keep you and your clankers safe w/ UV_MALWARE_CHECK=1 Charlie and team are the heroes we don’t deserve but definitely need!! Thank you Charlie & t... 原帖
  • 7 月 18 日 10:16，@w01fe，转发：RT @boazbaraktcs: I get this reaction too and it is unfortunate. AI's impact is so significant that it is important to involve many people… 原帖
  • 7 月 18 日 10:35，@nickbaumann_，转发：RT @dkundel: Beginning with the launch last week, GPT-5.6 Sol will be included in all Plus, Pro, Business, and Enterprise plans, at regular… 原帖
  • 7 月 18 日 10:56，@nickbaumann_，转发：RT @omooretweets: ChatGPT Sites is severely underrated - it is THE breakthrough in bringing vibe coding to the masses I prompted 5.6 to ma… 原帖
  • 7 月 18 日 11:04，@jxnlco，转发：RT @andrewtorkbaker: 150,000 people bought IMAX tickets to see The Odyssey the first day they were on sale. I was not one of them. So I bu… 原帖
  • 7 月 18 日 11:05，@jxnlco，原帖：Good use of tokens 原帖
  • 7 月 18 日 11:20，@JuhanaPeltomaa，原帖：the protagonist https://t.co/z3VW0i4AXI 原帖
  • 7 月 18 日 11:24，@jxnlco，转发：RT @MattHartman: After Venmo / Braintree were acquired by Paypal, they removed all the great apis from Venmo that let developers do interes… 原帖
  • 7 月 18 日 11:32，@jxnlco，原帖：Missed my flight yestuday and now I’m 72H sitting beside a man eating a tuna sandwich as we sit on the tarmac for an extra hour. 原帖
  • 7 月 18 日 11:54，@yong_zhengxin，转发：RT @yacineMTB: notice how they went from "they're 6 months behind the frontier" to "they're 6 months behind the unreleased frontier" 🤣🤣 原帖
  • 7 月 18 日 11:55，@andrewho03，原帖：I can’t walk around dodging San Francisco without dodging human feces and you think Chinese scientists want to return to China because of their hukou? 原帖
  • 7 月 18 日 11:56，@w01fe，转发：RT @tszzl: old news to many of you but this is a great meditation on the fragile world and the degree of abyss staring you are doing and wh… 原帖
  • 7 月 18 日 12:07，@andrewho03，转发：RT @suchenzang: what frustrates me the most about the "ccp is blind" takes from people who have never experienced the ccp in any form is th… 原帖
  • 7 月 18 日 12:13，@TheRohanVarma，原帖：I will accept your term sheet if you can get me tickets to see The Odyssey in 70mm 🙏🏾🫱🏼🫲🏾 原帖
  • 7 月 18 日 12:23，@andrewho03，转发：RT @suchenzang: just existing as an insult to the europoors without a worry for agi authoritarian gods enjoying the fruits of labor and in… 原帖
  • 7 月 18 日 12:38，@charliermarsh，原帖：Next Codex release gets incremental rendering for streamed Markdown in the TUI, hundreds of times faster for large outputs https://t.co/bbvDjkC0f9 原帖
  • 7 月 18 日 13:03，@npew，转发：RT @emollick: Computer use with Codex is really impressive: "I've never used Blender, download and install it using your computer control a… 原帖
  • 7 月 18 日 13:19，@JoanneShang，原帖：i think command + command (appshots in codex) has replaced command + k as my favorite shortcut 原帖
  • 7 月 18 日 13:20，@thsottiaux，原帖：What are you building this weekend? 原帖
  • 7 月 18 日 13:23，@andrew_n_carr，原帖：Gemini for google workspace enterprise vertex studio with advanced ultra pro account 原帖
  • 7 月 18 日 13:23，@steipete，转发：RT @thingsdev14: @steipete https://t.co/MlTsNg0MkG 原帖
  • 7 月 18 日 13:31，@ryanbrewer，转发：RT @kobex___: I lead Strategic Finance Foundations at @OpenAI, and I use Codex for everything. It all started with gaining leverage on our… 原帖
  • 7 月 18 日 13:42，@BorisMPower，转发：RT @thsottiaux: GPT-5.6 Sol confirmed to be an extremely good model 原帖
  • 7 月 18 日 13:42，@karinanguyen，转发：RT @bertgodel: Read the write up here and Karina's blog here: https://t.co/rgO5TXeNuW https://t.co/CrOuMciQWo 原帖
  • 7 月 18 日 13:42，@steipete，转发：RT @davemorin: OpenClaw for Human Rights 原帖
  • 7 月 18 日 13:50，@steipete，转发：RT @badlogicgames: one of these days i want peter to post something trueky absurd (this is not, iykyk) just to see the whole industry of AI… 原帖
  • 7 月 18 日 14:04，@giffmana，转发：RT @suchenzang: i was wondering when someone would finally call out how icky it is to see professors continue riding coattails of their stu… 原帖
  • 7 月 18 日 14:19，@sherwinwu，转发：RT @ChatGPTapp: Create something worth sharing with ChatGPT sites. Share your site with us for a chance to win a limited-edition swag box,… 原帖
  • 7 月 18 日 15:28，@ryanbrewer，原帖：Kyle has done some pretty crazy things with codex and spreadsheets at OpenAI 原帖
  • 7 月 18 日 15:43，@BorisMPower，原帖：Wow, such an exciting game - tune in! https://t.co/0ikRq1njmO 原帖
  • 7 月 18 日 16:10，@ericmitchellai，原帖：/gooooooooooooooooooooooal 原帖
  • 7 月 18 日 16:13，@jxnlco，转发：RT @thsottiaux: What are you building this weekend? 原帖
  • 7 月 18 日 16:17，@jeremyli__，原帖：Moonshot cooked 原帖
  • 7 月 18 日 16:17，@jeremyli__，转发：RT @andrewho03: https://t.co/v7RlDg1cTV 原帖
  • 7 月 18 日 16:20，@jxnlco，转发：RT @mitchellh: Doing stuff like this in the background lets me maximize my productive time when I actually have time to sit at a computer (… 原帖
  • 7 月 18 日 16:23，@andrewho03，原帖：Interestingly, GeneBench-Pro results seem to be highly correlated with prinzbench results. 原帖
  • 7 月 18 日 16:23，@BorisMPower，原帖：Such an incredibly fun game, I wish it were the finals https://t.co/2oxREuCVbR 原帖
  • 7 月 18 日 16:30，@ryanbrewer，转发：RT @jakeottiger: i scraped 10,000 codex posts, mapped their locations, and made them all searchable https://t.co/LurqL0IP2l in japan, a t… 原帖
  • 7 月 18 日 16:36，@thsottiaux，原帖：How can we make ChatGPT better. It’s already on your phone, but what do you wish would happen more or less of when you open the app. How can it bring ... 原帖
  • 7 月 18 日 16:40，@aidan_mclau，转发：RT @26thAndGuad: “Here’s Telemachus’ mentor. He’s named, uh, Mentor.” BRAVO NOLAN 原帖
  • 7 月 18 日 16:41，@jeremyli__，转发：RT @WillManidis: My friend Dean Ball has advanced an argument for the de facto protection of American frontier intelligence providers. Dea… 原帖
  • 7 月 18 日 16:42，@karinanguyen，原帖：insane match, england rlhf’d the life out of france today!! 原帖
  • 7 月 18 日 16:52，@jxnlco，原帖：is it safe to drive from washington dc to new york tn? 原帖
  • 7 月 18 日 16:58，@_rajanagarwal，原帖：hillclimbing https://t.co/8hyzsCSwvn 原帖
  • 7 月 18 日 17:09，@zhansheng，原帖：Online discourse on here is really bad. The amount of intentional misreading (“this is what Dean/OpenAI advocates”) and putting words in his mouth (“y... 原帖
  • 7 月 18 日 17:30，@andrewho03，转发：RT @GuiveAssadi: Andrew Ho is a VERY intelligent man and I take this result EXTREMELY seriously 原帖
  • 7 月 18 日 17:48，@gabrielchua，转发：RT @borvibe: Paris @OpenAI Build Week was lit!!!🇫🇷 Thanks amazing builders for coming ⚡️⚡️ thanks amazing host @42AI_ https://t.co/Lm3wljc… 原帖
  • 7 月 18 日 17:57，@gabrielchua，转发：RT @ChatGPTapp: Create something worth sharing with ChatGPT sites. Share your site with us for a chance to win a limited-edition swag box,… 原帖
  • 7 月 18 日 17:57，@andrewho03，转发：RT @teortaxesTex: This is a pretty compelling demonstration that Kimi K3 is both generally intelligent (on some level at least) and general… 原帖
  • 7 月 18 日 18:09，@nickbaumann_，原帖：The Odyssey lives up to the hype Doing whatever I can to see it on IMAX at the Metreon (If anyone's selling an extra ticket btw, please reply/DM) http... 原帖
  • 7 月 18 日 18:29，@jxnlco，原帖：Everyone talks about taste. But what about vision and courage. 原帖
  • 7 月 18 日 18:33，@andrew_n_carr，原帖：Movement is a fundamental aspect of the physical world we live in, and yet, it somehow slides into the background. Check out this example I care so de... 原帖
  • 7 月 18 日 18:52，@gabrielchua，转发：RT @YuYoshimuta: DJ コントローラーを Codex コントローラーへと魔改造するプロダクトで、Global Build Week Tokyo ハッカソン 1 位をいただきました！ @nukonuko さん、素敵なイベントを企画・運営してくださり、ありがとうございました。… 原帖
  • 7 月 18 日 18:58，@rapha_gl，原帖：if you’re an “accelerationist” ask yourself if that’s a technological project or a political one and if the latter, what re you accelerating towards 原帖
  • 7 月 18 日 19:07，@santa_kaus，转发：RT @thsottiaux: How can we make ChatGPT better. It’s already on your phone, but what do you wish would happen more or less of when you ope… 原帖
  • 7 月 18 日 19:55，@gabrielchua，转发：RT @phuctm97: Building a brand visual identity / guidelines with GPT 5.6 Sol Max is so damn fun. I feel like I'm a professional designer,… 原帖
  • 7 月 18 日 20:02，@jeremyli__，转发：RT @teortaxesTex: This is a pretty compelling demonstration that Kimi K3 is both generally intelligent (on some level at least) and general… 原帖
  • 7 月 18 日 20:08，@BarrAlexandra，原帖：frontier human data eval 原帖
  • 7 月 18 日 20:08，@danintheory，转发：RT @deanwball: I'd like to do two things: (1) tell you where I think I erred in my original post about Kimi and (2) set the record straight… 原帖
  • 7 月 18 日 20:23，@prd_008，转发：RT @thsottiaux: If you use ChatGPT Work. Looking to assemble a group of 50-100 to give early access to features and discuss how we can make… 原帖
  • 7 月 18 日 20:27，@gabrielchua，原帖：https://t.co/mdyVIC1Zsh 原帖
  • 7 月 18 日 20:41，@gabrielchua，转发：RT @gabrielchua: Looking for a weekend project? OpenAI Build Week closes in 60+ hours - plenty of time to build something useful with Code… 原帖
  • 7 月 18 日 20:41，@gabrielchua，原帖：👀 https://t.co/EWU8sV73BG 原帖
  • 7 月 18 日 20:46，@andrew_n_carr，原帖：Anyone know grok 4.5 posttrainbench score? 原帖
  • 7 月 18 日 21:08，@jxnlco，转发：RT @nickbaumann_: The Odyssey lives up to the hype Doing whatever I can to see it on IMAX at the Metreon (If anyone's selling an extra ti… 原帖
  • 7 月 18 日 21:25，@jxnlco，转发：RT @Da7_Tech: After using Sol heavily, I’m seriously considering upgrading to the $200 Codex plan. 原帖
  • 7 月 18 日 21:26，@jxnlco，转发：RT @billyjhowell: GPT sites are a much needed addition to Codex from @OpenAI. A little wonky. But I expect the product to improve fast. I… 原帖
  • 7 月 18 日 21:38，@_samirism，转发：RT @thsottiaux: ChatGPT Work is for ✅ Creating and hosting sites ✅ Managing your emails for you ✅ Summarizing mountains of documents ✅ Cre… 原帖
  • 7 月 18 日 21:42，@gabrielchua，转发：RT @thsottiaux: ChatGPT Work is for ✅ Creating and hosting sites ✅ Managing your emails for you ✅ Summarizing mountains of documents ✅ Cre… 原帖
  • 7 月 18 日 21:45，@gabrielchua，转发：RT @gabrielchua: https://t.co/mdyVIC1Zsh 原帖
  • 7 月 18 日 21:49，@nickbaumann_，转发：RT @thsottiaux: ChatGPT Work is for ✅ Creating and hosting sites ✅ Managing your emails for you ✅ Summarizing mountains of documents ✅ Cre… 原帖
  • 7 月 18 日 21:56，@jxnlco，转发：RT @thsottiaux: ChatGPT Work is for ✅ Creating and hosting sites ✅ Managing your emails for you ✅ Summarizing mountains of documents ✅ Cre… 原帖
  • 7 月 18 日 21:57，@jxnlco，转发：RT @thsottiaux: If you use ChatGPT Work. Looking to assemble a group of 50-100 to give early access to features and discuss how we can make… 原帖
  • 7 月 18 日 22:04，@neelajj，转发：RT @thsottiaux: ChatGPT Work is for ✅ Creating and hosting sites ✅ Managing your emails for you ✅ Summarizing mountains of documents ✅ Cre… 原帖
  • 7 月 18 日 22:06，@steipete，转发：RT @Pat_Erichsen: OpenClaw repo crossing 100k PRs is making Codex app highlight the references as hex color codes 💀 https://t.co/KQDaXYUOiW 原帖
  • 7 月 18 日 22:16，@andrewho03，原帖：The Chinese have many thoughts. 原帖
  • 7 月 18 日 22:23，@andrewho03，原帖：With regard to Chinese talent—it is especially relevant that AI and modern LLMs are quite new technology, so the ability of a given individual to cont... 原帖
  • 7 月 18 日 22:25，@yong_zhengxin，转发：RT @peterwildeford: The new era of cyberattacks -- HuggingFace reports being hacked by an AI and they defended with an AI! > "it was drive… 原帖
  • 7 月 18 日 23:00，@jxnlco，转发：RT @shubhampatilsd: @bilaltwovec https://t.co/uclBDLxJEI 原帖
  • 7 月 18 日 23:01，@jxnlco，转发：RT @siegerts: Codex 🏖️ Miami https://t.co/zJIaUflh8x 原帖
  • 7 月 18 日 23:55，@andrewho03，原帖：Yes, this makes no sense. If open source models proliferate, we will see relatively wider uptake of AI. Of course, one can construct rather involved a... 原帖
  • 7 月 19 日 00:43，@gabrielchua，转发：RT @Aman25m: Cafe Codex - Build Week Edition Pune was all about one thing. Building. Grateful for everyone who pushed ideas forward and m… 原帖
  • 7 月 19 日 01:41，@dylancablayan，转发：RT @athyuttamre: care is the moat 原帖
  • 7 月 19 日 02:27，@andrewho03，原帖：It seems straightforwardly good for the world that frontier model capabilities can be easily extracted and replicated elsewhere. Of course, it is not ... 原帖
  • 7 月 19 日 02:28，@gabrielchua，原帖：spotted at singapore codex community demo night https://t.co/RWvcFe4Ij4 原帖
  • 7 月 19 日 02:40，@andrewho03，原帖：I’m not really sure why there is some expectation that frontier labs must endlessly acquire companies and metastasize. Personally, I think that it wou... 原帖
  • 7 月 19 日 06:11，@gabrielchua，转发：RT @lucastohdev: Second day at OpenAI build week developer challenge, met some cool people like @gabrielchua who works at OpenAI and some o… 原帖
  • 7 月 19 日 06:50，@gabrielchua，转发：RT @nukonuko: Global Build Week Community Event - Tokyo という Codex を使ったハッカソンを無事開催することができました！#aimeetup お申し込み 159 人のうち最終的に参加予定 70 名、69 人が来てくだ… 原帖
  • 7 月 19 日 07:33，@SebastienBubeck，原帖：Talking about crazy results ... 原帖
  • 7 月 19 日 07:36，@ALupsasca，转发：RT @deanwball: Some observations on Kimi: 1. It's a very good model! I don't think its performance can be explained away by distillation… 原帖
  • 7 月 19 日 07:44，@reach_vb，原帖：you can just prompt things into existence 原帖
  • 7 月 19 日 07:46，@gabrielchua，转发：RT @eigen_moomin: you must strive to one day live in a society so safe and swagless that they fine you for jaywalking and it makes the nati… 原帖
  • 7 月 19 日 07:47，@BorisMPower，转发：RT @paulg: Trial lawyers are lobbying against self-driving cars because they're too safe. They need people to be killed and injured so that… 原帖

Anthropic

  • 7 月 18 日 09:16，@trq212，原帖：This was due to a heroic effort by many people at Anthropic working sometimes literally around the clock It was not at all clear that we'd be able to ... 原帖
  • 7 月 18 日 09:51，@RobertJBye，转发：RT @trq212: This was due to a heroic effort by many people at Anthropic working sometimes literally around the clock It was not at all cle… 原帖
  • 7 月 18 日 10:22，@RobertJBye，原帖：Love some morning tinkering on personal projects with Fable https://t.co/fJX1Q2WoFo 原帖
  • 7 月 18 日 11:42，@_sholtodouglas，原帖：interesting + nice test 原帖
  • 7 月 18 日 16:17，@jackclarkSF，原帖：Legendary game. 原帖

xAI

  • 7 月 18 日 08:32，@milichab，转发：RT @elonmusk: Try Grok Build https://t.co/qrq6Gn3HCS 原帖
  • 7 月 18 日 11:12，@milichab，原帖：Big updates next week :) 原帖
  • 7 月 18 日 11:48，@ypwang61，转发：RT @KyleLiang5: A simple loop has shown sign of life on not only pretraining perplexity, but also post-training reasoning tasks. Base CORE… 原帖
  • 7 月 18 日 16:57，@milichab，转发：RT @VineethSendil: Had a ton of fun shipping the /timeline toggle between experiments for the upcoming Grok models! We move super fluidly… 原帖
  • 7 月 18 日 21:10，@hyhieu226，转发：RT @chamath: The future is open source. We need to embrace it and get on with it. Imagine if America closed the door on open source. We… 原帖
  • 7 月 18 日 21:46，@sivil_taram，转发：RT @Daniel_Farinax: I asked Grok 4.5 to build macOS 27 with real Liquid Glass and native apps… It delivered in under 10 minutes. Kimi needs… 原帖

Google DeepMind

  • 7 月 18 日 08:52，@Majumdar_Ani，原帖：https://t.co/gQAP6hkhb0 原帖
  • 7 月 18 日 12:00，@suchenzang，原帖：what frustrates me the most about the "ccp is blind" takes from people who have never experienced the ccp in any form is that arguably the entire prob... 原帖
  • 7 月 18 日 16:27，@JeffDean，转发：RT @giffmana: It's kinda fun that "pod" is becoming de facto industry standard for "bunch of highly connected accelerators" AFAIK "TPU Pod… 原帖

Meta AI

  • 7 月 18 日 11:40，@baaadas，原帖：It’s not even the semi-finals; Tuchel parked the bus in UEFA and got punished by it as well. To be honest, England might have gone home earlier if it ... 原帖
  • 7 月 19 日 03:10，@HeinrichKuttler，原帖：More math progress. 原帖

Thinking Machines Lab

  • 7 月 18 日 08:35，@shizhediao，转发：RT @hendrydong: “More on-policy” is NOT always better! ReOPD reveals a two-sided prefix shift: ① student occupancy mismatch ② teacher reli… 原帖
  • 7 月 18 日 10:41，@x_hanyu，转发：RT @arcprize: Inkling from @thinkymachines on ARC-AGI (Verified) - ARC-AGI-2: 36.5%, $0.64/task - ARC-AGI-1: 79.5%, $0.30/task As of toda… 原帖
  • 7 月 18 日 10:55，@ziqiao_ma，转发：RT @ArtificialAnlys: Thinking Machines' newly released 975B parameter open weights model, Inkling, supports transcription and ranks second… 原帖
  • 7 月 18 日 14:51，@SonglinYang4，转发：RT @guohao_li: why can the china labs build glm-5.2, kimi k3, and many more to come? it is because of the openness. not just the open weigh… 原帖
  • 7 月 18 日 14:51，@SonglinYang4，转发：RT @gneubig: Agree with this: US labs are underestimating the power and vitality of open research. This is one of my favorite essays and I… 原帖
  • 7 月 18 日 15:41，@ekzhang1，原帖：as I was reading this: y’know it’s really not easy to change a system once it has momentum this system was built years ago for one use case (10-20 t... 原帖
  • 7 月 18 日 15:42，@SonglinYang4，转发：RT @MaziyarPanahi: Yesterday I played Thinking Machines' Inkling a doctor's visit. Today I gave it her lab report and started asking questi… 原帖
  • 7 月 18 日 17:47，@SonglinYang4，转发：RT @zhang_benita: As Kimi K3 launches, check out our earlier podcast interview with Yang Zhilin, founder of Kimi. We recorded it right afte… 原帖
  • 7 月 18 日 17:58，@SonglinYang4，原帖：That matches my intuition exactly 原帖
  • 7 月 18 日 19:04，@x_hanyu，原帖：Interesting post — definitely expanded how I think about open weights. Open sourcing used to feel like a good way to get attention when the model wasn... 原帖
  • 7 月 18 日 19:49，@Haoxiang__Wang，转发：RT @natolambert: Inkling’s humility at launch may have made it underrated. The previewed small model will wake a lot of people up (if it c… 原帖
  • 7 月 18 日 20:07，@regulargio，原帖：Pretty cool work! 原帖
  • 7 月 18 日 21:21，@regulargio，原帖：Pretty unique / original use case!!! 原帖
  • 7 月 18 日 23:02，@ekzhang1，原帖：it’s been a good year. maybe some of the most significant events in my life happened :) we must believe we are on the right path when we are grateful ... 原帖
  • 7 月 19 日 01:22，@SonglinYang4，转发：RT @Alibaba_Qwen: Qwen3.8 is launching and going open-weight soon!🌐 With a massive 2.4T parameters, this model is continuously evolving. W… 原帖
  • 7 月 19 日 01:35，@SonglinYang4，转发：RT @yzhang_cs: Open source is thriving! and now every open-weight model >2T adopts GDN/KDA in FLA 👀 原帖
  • 7 月 19 日 01:35，@SonglinYang4，原帖：lol 原帖

Cursor

  • 7 月 19 日 06:07，@benln，原帖：Tip: never use the green Open to Work banner on LinkedIn Negative signal for hiring managers and recruiters 原帖

Alibaba Qwen

  • 7 月 19 日 01:29，@Alibaba_Qwen，原帖：Qwen3.8 is launching and going open-weight soon!🌐 With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the mos... 原帖

Cognition

  • 7 月 18 日 20:34，@mattbergland，原帖：building personal software for bachelor party games this is 2026 https://t.co/3h8A5jEY5x 原帖
  • 7 月 18 日 21:33，@dabit3，转发：RT @pingToven: and so it begins https://t.co/XZ1IeDAWgz 原帖
  • 7 月 19 日 07:03，@jhanikhil，原帖：jobs are exactly like dating with absolutely no differences whatsoever 原帖

Moonshot/Kimi

  • 7 月 18 日 08:54，@Kimi_Moonshot，转发：RT @togethercompute: We analyzed Kimi K3 vs. Claude Fable 5 for software engineering tasks using DeepSWE. Kimi K3 gets you the same perfor… 原帖
  • 7 月 18 日 21:10，@Kimi_Moonshot，转发：RT @ValsAI: Kimi K3 is the #2 overall model on our in-house Vibe Code Bench at 85.0%. VCB tests a model's ability to go from zero-to-one;… 原帖
  • 7 月 19 日 07:52，@Kimi_Moonshot，原帖：Kimi K3 has received far more love than we expected, and our GPUs are feeling it. Over the past 48 hours, demand has pushed close to the limits of our... 原帖
  • 7 月 19 日 07:58，@crystalsssup，转发：RT @Kimi_Moonshot: Kimi K3 has received far more love than we expected, and our GPUs are feeling it. Over the past 48 hours, demand has pu… 原帖

ByteDance Seed

  • 7 月 18 日 14:23，@GenAI_is_real，原帖：A 38-minute multi-speaker meeting turned into a fully diarized, timestamped transcript in 49 seconds on a single H100. At concurrency 16, that's 97.5 ... 原帖

AMI Labs / ex-Meta AI

  • 7 月 18 日 13:17，@ylecun，转发：RT @KenRoth: With Trump deeply unpopular, Republicans face a possible drubbing in congressional races in November. So Trump seems intent on… 原帖
  • 7 月 18 日 13:26，@ylecun，转发：RT @BrianRoemmele: Kimi K3 is a gut punch to Anthropic. Its over. A 2.8T open-weight-bound Chinese model just took the WebDev human-prefer… 原帖
  • 7 月 18 日 13:37，@ylecun，转发：RT @IntuitMachine: https://t.co/CHEk211f7Y 原帖
  • 7 月 18 日 13:40，@ylecun，转发：RT @RpsAgainstTrump: Document Declassified by Trump Reveals Russia Tried To Help Him Win the 2020 Election https://t.co/D5BZq6W5QZ 原帖
  • 7 月 18 日 13:42，@ylecun，转发：RT @YourAnonNews: Trump crying about rigged elections, but remember when he was literally recorded trying to rig an election? How the fuck… 原帖
  • 7 月 18 日 13:43，@ylecun，转发：RT @atrupar: to hear Trump tell it, American elections were completely corrupted while he was president and he was helpless to stop it, the… 原帖
  • 7 月 18 日 13:51，@ylecun，转发：RT @MAGALieTracker: BREAKING: The last month has cemented Trump's status as the most corrupt president in American history. It's not even c… 原帖
  • 7 月 18 日 13:51，@ylecun，转发：RT @RpsAgainstTrump: Here’s a montage of Trump saying over the past five months that the war with Iran is about to end and that the Iranian… 原帖
  • 7 月 18 日 13:53，@ylecun，转发：RT @Dan_Jeffries1: You either believe in centralization of power or broad diffusion of power. In one worldview, there are some bad people… 原帖
  • 7 月 18 日 13:54，@ylecun，转发：RT @martin_casado: "Open-weight models are inherently decelerationist" .... this is a grossly incorrect statement with no supporting argu… 原帖
  • 7 月 18 日 13:54，@ylecun，转发：RT @KenRoth: "Trump called for fixing vulnerabilities in electronic voting system" but he "has spent…his second term dismantling the electi… 原帖
  • 7 月 18 日 23:01，@ylecun，转发：RT @chamath: The future is open source. We need to embrace it and get on with it. Imagine if America closed the door on open source. We… 原帖
  • 7 月 18 日 23:52，@ylecun，转发：RT @Kasparov63: "Russia hoax" is as strong an indicator of a being gullible idiot or partisan propagandist as "vaccines are bad", and both… 原帖
  • 7 月 19 日 00:20，@ylecun，转发：RT @mjfree: 🚨🚨🚨 BREAKING: In a fully corrupt & desperate move, Trump is forcing the Postmaster General to REFUSE delivery of mail-in ballot… 原帖
  • 7 月 19 日 02:56，@ylecun，转发：RT @MelMitchell1: In the off chance than anyone cares what I think on this topic: (1) the open-source software movement has been enormous… 原帖
  • 7 月 19 日 04:23，@ylecun，转发：RT @ylecun: @deanwball The "ungovernability" (and openness) of Linux and the Internet is precisely what has made their success. The same w… 原帖
  • 7 月 19 日 04:25，@ylecun，转发：RT @martin_casado: Do you know what open weights model actually decelerate? The formation and power of oligopolies. You know, those thin… 原帖
  • 7 月 19 日 06:50，@ylecun，转发：RT @Suhail: They distilled humanity. The hypocrisy is unreal. The best gift is to give it away, at the lowest marginal costs, to let us p… 原帖
  • 7 月 19 日 06:54，@ylecun，转发：RT @Dan_Jeffries1: @hosseeb @deanwball Dumping? Come the fuck on. Stop trying to use clever language to make your points and make them hone… 原帖
  • 7 月 19 日 07:03，@ylecun，转发：RT @ylecun: @hosseeb @deanwball Soooo, releasing Linux was dumping? Apache, MySQL, PHP? HTTP, TCP/IP, OpenSSL, OpenSSH? Libjpeg, VLC? Th… 原帖

References

  1. 1Qwen 官方：Qwen3.8 开放权重预告
  2. 2Kimi 官方：K3 需求与会员调整
  3. 3Charlie Marsh：Codex TUI 增量渲染
  4. 4Vaibhav Srivastav：ChatGPT Work 托管浏览器
  5. 5Chayenne Zhao：H100 上的 MOSS-TD ASR 服务
  6. 6Boris Power：转发 GPT-5.6 Sol 评价
  7. 7Aidan McLaughlin：窗口内帖子
  8. 8Andrew Ho：开放模型扩散
  9. 9Andrew Ho：实验室收购与扩张
  10. 10Sholto Douglas：窗口内短评
  11. 11Thariq：Fable 团队协作
  12. 12Hanyu Xue：Inkling ARC-AGI 转发
  13. 13Martin Ziqiao Ma：Inkling 开放权重模型转发
  14. 14Hanyu Xue：开放权重策略讨论
  15. 15Shizhe Diao：ReOPD 与 on-policy 讨论
  16. 16Eric Zhang：系统从 10 至 20 TPS 扩展到 1000 倍负载
  17. 17Songlin Yang：开放性与中国实验室
  18. 18Songlin Yang：开放研究活力
  19. 19Kimi 官方：转发 DeepSWE 分析
  20. 20Jeff Dean：TPU Pod 与加速器集群讨论
  21. 21Andrew Milich：下周更新预告
  22. 22Yiping Wang：pretraining 与 post-training 讨论
  23. 23Jason Wolfe：AI 讨论需要更多参与者
  24. 24Songlin Yang：转发 Qwen3.8 预告

Related content

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