八月十八日

Google VP Josh Woodward reports that Gemini's revamped Workspace tools should enter testing within one to two weeks, while the team implements a finished design for Projects and continues frontend work after completing the corresponding backend. He also says Gemini now supports 49 connectors, Gemini 3.7 Flash improved tool calling, two other requested items are ramping after completion, and the largest over-triggering bugs have been fixed. The update shows the assistant layer being built around dependable tool use, persistent project context, and a widening set of external services rather than model quality alone.

Google VP Josh Woodward 表示，Gemini 改版后的 Workspace tools 预计将在一到两周内进入测试；Projects 的设计已经完成并进入实现阶段，另一个功能也在 backend 完成后开始 frontend 工作。他还提到 Gemini 目前支持 49 个 connectors，Gemini 3.7 Flash 已改善 tool calling，另外两个已完成项目正在逐步放量，最主要的误触发问题也已经修复。这份更新显示，assistant layer 的建设重点不再只是模型质量，而是可靠的工具调用、持久的项目 context，以及不断扩大的外部服务连接能力。

Swyx highlights Trajectory's work on continual learning, particularly its treatment of the remaining data problems. He says Ronak Pradeep explained why GRPO was insufficient, why the team moved on-policy, and how that shift exposed further issues that then had to be fixed. The useful signal is that continual learning is not a single algorithmic switch: ambitious systems have to manage the data distribution and operational complications created by learning from their own evolving behavior.

Swyx 重点介绍了 Trajectory 在 continual learning 上的工作，尤其是团队如何处理尚未解决的数据问题。他转述 Ronak Pradeep 的说明：GRPO 并不足够，团队因此转向 on-policy，随后又必须处理这种变化带来的一系列新问题。这里的重要信号是，continual learning 并不是切换一个算法就能完成；有野心的系统还必须管理由自身不断变化的行为所产生的数据分布与运行复杂性。

Meta senior director of AI Madhu Guru lays out a practical sequence for building evals: start with a workflow the team understands deeply, inspect the actual prompt traces and desired responses at each step, encode known failures such as messy tool results or missing context, then make the eval repeatable and automatic. The final requirement is maintenance, because the suite must continue to reflect live traffic as user behavior changes. This turns evaluation from a static launch gate into an operating system for measuring and improving an AI product over time.

Meta AI senior director Madhu Guru 给出了一套实用的 eval 构建顺序：先选择团队非常熟悉的 workflow，检查真实的 prompt traces 以及每一步和最终结果应有的表现，再把混乱的工具返回、context 缺失等已知失败编码进测试，最后让 eval 可以重复、自动运行。更长期的要求是持续维护，因为随着用户行为变化，测试集必须始终映射 live traffic。这种方法把 evaluation 从一次性的上线门槛，变成长期衡量和改进 AI 产品的运行系统。

Anthropic's Thariq says recent procedural-art, video-editing, and 3D-game demonstrations have shifted his view toward LLM coding models outperforming diffusion models across more creative tasks. His reason is structural: code is easier to edit, steer, and export into existing tools. His invitation to use Claude Code's `/design` command points toward a creative workflow where the model produces an inspectable and manipulable artifact, not only a finished generation.

Anthropic 的 Thariq 表示，近期 procedural art、video editing 和 3D game demos 让他更倾向于认为，LLM coding models 会在更多创意任务上胜过 diffusion models。他给出的原因是结构性的：代码更容易编辑、引导，也更容易导出到现有工具中。他同时邀请用户尝试 Claude Code 的 `/design` 命令，这指向一种新的创意 workflow：模型交付的不只是最终生成结果，而是一个可以检查和直接修改的 artifact。

Replit CEO Amjad Masad points to a team achieving AI-like growth without putting AI in its pitch, while operating with what he says would otherwise require ten times the headcount. In a separate security observation, he argues that vulnerability scanning is not sufficient and should be paired with penetration testing that actively tries to break the software. Together, the posts capture a useful production discipline: AI can radically compress organizational labor, but the resulting output still needs adversarial verification rather than passive inspection alone.

Replit CEO Amjad Masad 提到，一个团队虽然没有在 pitch 中强调 AI，却实现了 AI 式增长；按他的判断，如果不是深度采用 AI，这家公司需要十倍的员工规模。在另一条安全观察中，他强调 vulnerability scanning 并不足够，还需要通过 penetration testing 主动尝试攻破软件。两条信息共同指向一项重要的生产纪律：AI 可以大幅压缩组织所需的人力，但产出仍然需要对抗式验证，而不能只依赖被动检查。

Vercel CEO Guillermo Rauch says repositories can now be hosted in Cursor Origin and deployed from there to Vercel, while Cursor Origin itself runs on Vercel. Beyond the recursive architecture, the consequential product change is that an AI coding environment is moving upstream into source hosting and downstream into deployment. That creates a more continuous path from an agent's code changes to running software, while making the reliability and governance of the integrated workflow increasingly important.

Vercel CEO Guillermo Rauch 表示，repository 现在可以托管在 Cursor Origin，并从 Cursor Origin 直接部署到 Vercel，而 Cursor Origin 本身也运行在 Vercel 上。除了这种递归式架构，更重要的产品变化是：AI coding environment 正在向上游进入 source hosting，向下游延伸至 deployment。这让 agent 从修改代码到交付运行中软件的路径更加连续，同时也会让整个集成 workflow 的可靠性与治理变得更重要。

Box CEO Aaron Levie argues that AI's demand makes information valuable in almost any form and that organizational intelligence increasingly belongs on the balance sheet as an asset. FPV Ventures partner Nikunj Kothari approaches differentiation from another direction: after mocking the claim that every layer of AI lacks a moat, he argues that brand marketing will become a major asset, especially if agents become the primary users of many products and companies must still win scarce human attention. Taken together, their posts suggest that as models, harnesses, and infrastructure face fast imitation, proprietary information and a consistently articulated company identity may carry more of the durable value.

Box CEO Aaron Levie 认为，AI 对数据的强烈需求正在让几乎任何形态的信息都变得有价值，组织 intelligence 也越来越应该作为资产出现在 balance sheet 上。FPV Ventures partner Nikunj Kothari 则从另一个方向讨论差异化：他先讽刺了“AI 每一层都没有 moat”的说法，随后提出 brand marketing 将成为重要资产，尤其是在 agent 成为许多产品主要用户、公司却仍需争夺稀缺人类注意力的情况下。合并来看，两人的观点意味着，当模型、harness 和基础设施都面临快速复制时，专有信息与长期一致的公司表达可能承载更多持久价值。

Sarah Guo and Elad Gil challenge the assumption that many new AI companies can reach trillion-dollar scale within three to five years. Gil separates eventual market size from the speed required to reach roughly $50 billion to $100 billion in revenue, while Guo argues that outcome-based pricing can make application markets much larger than seat-based comparisons imply. They also propose that founders revisit exit decisions every six months as capabilities and costs move quickly, asking whether their companies actually capture the value created by that secular change.

Sarah Guo 和 Elad Gil 质疑了“未来三到五年会出现许多新的万亿美元 AI 公司”这一假设。Gil 区分了长期市场规模与在短期内达到约 500 亿至 1,000 亿美元收入所需的增长速度；Guo 则指出，按结果收费可能让 application 市场远大于按 seat 估算的传统类比。他们还建议 founder 每六个月重新审视一次退出决策，因为能力与成本变化极快，真正需要追问的是公司是否捕获了这轮长期变化创造的价值。

The conversation then moves from company valuation to resource allocation. Compute constraints can reinforce an oligopoly among major labs, while scarce token budgets push organizations to measure "return on invested tokens" and concentrate resources on the people and projects with the highest expected impact. Guo is cautious about recurring predictions of recursive self-improvement being eighteen months away, and both hosts discuss how such compressed timelines can distort researchers' work and life choices. Their closing policy argument is similarly framed as allocation under uncertainty: safeguards matter, but regulation that counts risk without weighing benefits can entrench incumbents, slow health and productivity gains, and repeat mistakes they see in energy and biotech.

随后，讨论从公司估值转向资源分配。Compute 约束可能强化主要 labs 之间的 oligopoly，而稀缺的 token budget 会推动组织衡量“return on invested tokens”，并把资源集中到预期影响最大的人员与项目。Guo 对不断重复出现的“recursive self-improvement 距离我们还有十八个月”预测保持谨慎，两位主持人也讨论了这种压缩时间线如何扭曲研究人员对工作与生活的选择。最后的政策讨论同样被放在不确定性下的资源配置框架中：safeguards 很重要，但如果监管只计算风险、不衡量收益，就可能巩固 incumbent、拖慢医疗与生产力进步，并重复他们认为已经在能源和 biotech 中出现的错误。