🌐 双语
Archive

AI Builders
Digest

2026-11-18 13 builders · 33 tweets · 1 podcasts · 0 blogs

🔥 热点话题

OpenAI 发布 6.2 Sol 与 Astra:高效但体验仍需打磨OpenAI Launches 6.2 Sol and Astra: Highly Capable Yet Still Rough Around the Edges

OpenAI CEO Sam Altman 宣布 6.2 Sol 与 Astra 正式上线。这些模型在智能和能力上表现非凡,且 token 效率极高,尤其是 6.2 Sol ultrafast 版本让人印象深刻。他同时承认,模型在“好用程度”和交互体验上仍有不足,团队会优先改进。

Codex & ChatGPT 负责人 Thibault Sottiaux 补充,早期访问用户与大众反馈出现落差,是因为最后 3-4 天做了大量改动,而评测视频大多基于更早版本。他强调模型在速度、成本和任务完成效率上提升显著,并推荐 Pro 用户尝试 Sol ultrafast。

Every CEO Dan Shipper 团队给出的评分显示,这是继 Opus 5 以来最接近“里程碑”的一次前沿模型发布。Latent Space 主持人 Swyx 虽批评 Astra 广告决策失误,但承认模型本身是他用过最好的工作流自动化模型,已在内容发布管道中稳定运行。
OpenAI CEO Sam Altman announced the release of 6.2 Sol and Astra. The models deliver extraordinary intelligence and capability while being highly token-efficient; he particularly highlights 6.2 Sol ultrafast. At the same time he acknowledged that “making the models nice to work with” still needs work and will be prioritized.

Codex & ChatGPT lead Thibault Sottiaux explained the gap between influencer and public sentiment: early access was granted weeks ago, then major changes landed in the final 3-4 days, yet many review videos were filmed on the older builds. He stresses the models’ efficiency in speed, cost, and task completion, and recommends Pro users try Sol ultrafast.

Every CEO Dan Shipper’s team ratings call it the closest frontier-model release to Opus 5 in recent memory. Latent Space host Swyx criticized the Astra ad decision but called the model itself the best workflow-automation system he has used, already powering content pipelines with strong ROI.
查看原文 →查看原文 →查看原文 →查看原文 →

Black Forest Labs:从图像生成到视觉智能与机器人Black Forest Labs: From Image Generation to Visual Intelligence and Robotics

The Takeaway:Black Forest Labs 正把扩散模型从图像生成扩展到完整视觉智能与机器人控制,目标是在 2027 年实现单次示范学习。

Andreas Blattmann(Andi)是 latent diffusion 的共同发明人、Black Forest Labs 联合创始人;Dustin Podell 从软件工程师一路成长为 SDXL、Flux 系列的研究负责人。两人离开 Stability AI 后在德国黑森林地区创立 BFL,目前已累计融资 4.5 亿美元,收入突破 1 亿美元。

他们把图像、视频、音频和动作预测统一到 Flux.3 多模态模型中,强调视频建模是理解世界的核心路径。模型通过 2D 视频数据即可获得隐式 3D 与物理直觉,无需显式 3D 表示。Flux Context 最初只是 Dustin 的一个 side project,用极少量高质量关系数据训练数小时后就展现出通用编辑能力,彻底改变了“抽卡式”生成流程。

在机器人方向,7B 的 Flux Action 模型仅需数小时数据即可微调到可用水平,目标是实现“给机器人看一次动作,它就能复现”的超级人类视觉智能。Andi 说:“You should be showing a robot an action one time, and then it should be able to do this, and this would be truly a superhuman intelligence.”
The Takeaway: Black Forest Labs is extending diffusion models beyond image generation into full visual intelligence and robotics, with a clear aim of single-shot learning by 2027.

Andreas Blattmann (Andi), co-inventor of latent diffusion and co-founder of Black Forest Labs, together with Dustin Podell—who rose from software engineer to research lead behind SDXL and the Flux series—left Stability AI to start BFL in Germany’s Black Forest region. The company has raised $450 million and already surpassed $100 million in revenue.

They unified image, video, audio and action prediction inside the Flux.3 multimodal model, arguing that video modeling is the central route to world understanding. Models trained only on 2D video acquire implicit 3D and physics intuition without any explicit 3D representation. Flux Context began as Dustin’s side project; after a few hours of training on a tiny high-quality relationship dataset it suddenly generalized, ending the “slot-machine” era of image editing.

On robotics, the 7B Flux Action model reaches usable performance after only hours of fine-tuning data. The explicit goal is single-shot learning: show a robot an action once and it can perform it. As Andi puts it: “You should be showing a robot an action one time, and then it should be able to do this, and this would be truly a superhuman intelligence.”
查看原文 →

🛠️ 开发者工具与技巧

Devin Bots:真正可用的多 Agent 团队Devin Bots: Multi-Agent Teams Done Right

Latent Space 主持人 Swyx 详细介绍了 Cognition 的 Devin Bots。它们是常驻、主动(但受 ROI 约束)、拥有结果而非任务、可管理子 Agent 并自我复盘的系统,覆盖 Slack、短信、电话等全渠道,并通过交互式 artifact(HTML)而非长聊天汇报。

Swyx 已把 AEO、QA、跨平台移植、个人行政和视频剪辑全部交给 Bots。他强调:我们还没有 AGI,不要承诺“AI 员工”,真正需要的是比单任务更高一层的“Agent Manager”——拥有问题域、主动发现问题并带来解决方案的角色。这也是他衡量人类员工“是否有 agency”的同一标准。
Latent Space host Swyx unpacked Cognition’s Devin Bots: always-on, proactive (ROI-tuned), outcome-owning rather than task-owning agents that manage sub-agents, review themselves, and are reachable via Slack, text, phone or any channel. Reporting happens through interactive HTML artifacts instead of endless chat logs.

Swyx has already handed over AEO, QA, cross-platform porting, personal admin and video editing. His core insight: we do not yet have AGI, so stop promising “AI employees.” What works is the next abstraction layer above single tasks—agent managers that own a problem area, surface issues and bring solutions. It is the same filter he uses when hiring humans.
查看原文 →查看原文 →

Replit 一键启动 Agent SwarmReplit Makes Agent Swarms One Prompt Away

Replit CEO Amjad Masad 宣布:Agent Swarm 是 AI 的下一个扩展维度。过去我们靠更大模型和更长思考时间扩展智能,现在可以并行运行大量 Agent,让它们互相挑战、探索不同路径并协作解决单 Agent 难以完成的问题。

OpenAI 已在数学突破上证明了这一路线。此前自己搭建编排基础设施门槛很高,从今天起在 Replit 上只需一条 prompt 即可启动 Swarm。Amjad 期待看到集体机器智能被普通人编程后会发生什么。
Replit CEO Amjad Masad declared agent swarms the next scaling frontier. After years of scaling via larger models and longer thinking time, the new axis is running many agents in parallel so they can explore different approaches, challenge one another and solve problems no single agent could easily crack.

OpenAI has already shown the power of this approach in mathematics. Until now the orchestration overhead was prohibitive; starting today on Replit a swarm is one prompt away. Amjad is excited to see what becomes possible when collective machine intelligence is programmable by anyone.
查看原文 →查看原文 →

Gemini 免费任务助手与 Google Workspace 深度集成Gemini Free Task Automation + Deep Workspace Integration

Google Labs VP Josh Woodward 宣布 Gemini 在美国、日本、韩国、加拿大、澳大利亚和印度六国免费提供任务执行能力,运行在新的 Gemini 4 Hydrogen 模型上。新增功能包括:通过 Google Duplex 打电话给本地商家、Google Pay 结账、RCS 短信交互、转发邮件给 Gemini、用密码管理器安全登录,以及更可靠的 Workspace 工具调用(起草/发送邮件、创建编辑 Docs/Sheets、批量添加日历事件)。
Google Labs VP Josh Woodward announced that Gemini can now complete tasks free of charge for users in the US, Japan, Korea, Canada, Australia and India, powered by the new Gemini 4 Hydrogen model. New capabilities include calling local businesses via Google Duplex, paying with Google Pay, texting via RCS, forwarding emails, password-manager sign-in, and a major Workspace upgrade that lets users draft or send email, create/edit Docs and Sheets, and add multiple Calendar events without leaving Gemini.
查看原文 →

Vercel:自优化软件与网络安全模型Vercel: Self-Optimizing Software and Cyber-Capable Models

Vercel CEO Guillermo Rauch 指出 agent-browser vitals 是网页性能的巨大进步——软件开始自我优化。他同时称赞 OpenAI 最新模型在网络安全能力上达到世界领先水平,并提到 Vercel 的 DeepsecBench 是隐蔽基准,没有任何实验室能提前看到代码进行训练。他还推荐了 fx 工具。
Vercel CEO Guillermo Rauch called agent-browser vitals a major leap for web performance—self-optimizing software is here. He also highlighted that OpenAI just released the most cybersecurity-capable models in the world, noting that Vercel’s DeepsecBench is a stealth benchmark no lab has been able to train on. He additionally recommended the fx tool.
查看原文 →查看原文 →查看原文 →

💰 创业成功案例

FirstMark 领投 Remanence Labs 3000 万美元种子轮FirstMark Leads $30M Seed in Remanence Labs for Continual Learning

FirstMark VC Matt Turck 宣布与 EQT 共同领投 Remanence Labs 3000 万美元种子轮,并与创始人合作。公司聚焦持续学习(continual learning),让智能能够不断累积而非一次性训练。Turck 同时提到 Daytona 推出 macOS 沙箱,专为计算机使用 Agent 以及 Apple 应用的构建与测试设计。
FirstMark VC Matt Turck announced a $30M seed co-led with EQT into Remanence Labs, partnering with the founders on continual learning—intelligence that compounds rather than being trained once. He also highlighted Daytona’s new macOS sandboxes, the first infrastructure tailored for computer-use agents on Macs and for building/testing Apple apps.
查看原文 →查看原文 →

🌍 其他动态

模型品味与短生命周期观察Model Vibes and Short Shelf-Life Observations

Practical AI 博主 Peter Yang 指出,除了少数特例(如 Opus 5.5),大多数模型的“保质期”极短,几周后大家就转向下一个。不必对单一模型投入过多情绪,关键是让自己的技能能跨模型迁移。他同时测试 Astra 6.2 在网站设计上似乎已超过 Fable。

FPV Ventures 合伙人 Nikunj Kothari 观察到模型“氛围”在 X 上剧烈摇摆:前几个月 OpenAI 占优,现在 Anthropic 5.5 与 Gemini Carbon 获得更多赞誉,期待 Grok 带来惊喜。
Practical AI creator Peter Yang observed that most models (except Opus 5.5 and a few specials) have extremely short shelf lives—weeks later everyone moves on. There is little point getting emotional about any single release; the real skill is making one’s workflows portable across models. He also found Astra 6.2 already better than Fable at website design in early tests.

FPV Ventures partner Nikunj Kothari noted the wild yo-yo of model vibes on X: OpenAI held the mantle for months, now Anthropic 5.5 and Gemini Carbon are collecting the accolades, and it is Grok’s turn to surprise.
查看原文 →查看原文 →查看原文 →