AI Builders Digest | 2026-09-17
X / TWITTER
Google VP Josh Woodward
Josh Woodward highlighted two Gemini Notebook upgrades aimed at students: live spoken Q&A with class materials in roughly 100 languages, and automatic capture of recorded lectures and audio notes into a chosen notebook. Google is also continuing its free AI Plan offer for university students across more than 140 countries.
Josh Woodward 重点介绍了 Gemini Notebook 面向学生的两项升级：可用约 100 种语言与课程材料进行实时语音问答，以及把录制的讲座和语音笔记自动保存到指定 notebook。Google 还在 140 多个国家继续向大学生提供免费的 AI Plan。
https://x.com/joshwoodward/status/2099921866014306633
Claude Code team member Thariq
Thariq argues that MCP has become better than CLI for most integrations because models now call tools more reliably, tools can be loaded only when needed, and MCP can operate statelessly. For workflows that need composition or filtering, he recommends exposing query-style parameters directly in MCP tools.
Claude Code 团队成员 Thariq 认为，对大多数集成而言，MCP 已经优于 CLI：模型的 tool calling 更可靠，工具可以按需加载，MCP 也已支持 stateless 模式。对于需要组合或过滤数据的场景，他建议直接在 MCP 工具中提供 query 一类参数。
https://x.com/trq212/status/2099958388230873165
Replit CEO Amjad Masad
Amjad Masad questioned whether a general probability-producing model is necessary when the output domain is known in advance. His alternative is simpler: train a model to emit log probabilities over a fixed enum, which may be more direct and efficient for constrained judgment tasks.
Replit CEO Amjad Masad 质疑：如果输出空间事先已知，是否真的需要一个通用的概率输出模型。他提出更直接的做法：让模型针对固定 enum 输出 log probabilities，这可能更适合受约束的判断任务，也更高效。
https://x.com/amasad/status/2100056178705514703
Vercel CEO Guillermo Rauch
Guillermo Rauch formally launched Vercel Labs as Vercel's public research and experimentation arm, including disclosure of supported projects, active research, and experiments that failed. He also predicts WebAssembly will become central to the web as more code goes native, citing Safari 27's JSPI support, and argues that products should expose multi-model choice instead of hiding it from users.
Vercel CEO Guillermo Rauch 正式推出 Vercel Labs，作为 Vercel 面向公众的研究与实验部门，公开其支持的项目、研究方向以及未成功的实验。他还判断，随着更多代码走向 native，WebAssembly 将成为未来 Web 的核心，并以 Safari 27 支持 JSPI 为例；同时他主张产品应明确提供 multi-model 选择，而不是替用户隐藏模型差异。
https://x.com/rauchg/status/2099911447598059812
https://x.com/rauchg/status/2099974859023683975
https://x.com/rauchg/status/2099905740505055680
Box CEO Aaron Levie
Aaron Levie sees the gap between model capability and real enterprise workflows as the applied AI layer's opportunity. Winning products must connect intelligence to systems and data, redesign processes, preserve human oversight, run domain-specific evals, and handle security, governance, and change management; stronger models expand this need by making more complex workflows automatable.
Box CEO Aaron Levie 认为，模型能力与企业真实工作流之间的巨大鸿沟，正是 applied AI layer 的机会。真正有效的产品必须连接系统与数据、重构流程、保留 human-in-the-loop、执行领域专属 eval，并处理安全、治理和组织变革；模型越强，可自动化的任务越复杂，这一层反而越重要。
https://x.com/levie/status/2099976021311398230
Y Combinator CEO Garry Tan
Garry Tan reported that Capy, used with GStack and GBrain on outstanding issues and pull requests, completed work in roughly half the time required by raw Codex or Claude Code while using the same frontier models. The signal is that workflow orchestration and task scaffolding can create substantial gains even when underlying model quality is held constant.
Y Combinator CEO Garry Tan 表示，他用 Capy 配合 GStack 和 GBrain 处理遗留 issue 与 PR 时，在使用相同 frontier models 的情况下，耗时约为直接使用 Codex 或 Claude Code 的一半。这说明即使底层模型不变，工作流编排与任务脚手架仍能带来显著效率增益。
https://x.com/garrytan/status/2099964487667454097
FPV Ventures partner Nikunj Kothari
Nikunj Kothari warns founders not to assume the next funding round will arrive automatically. Capital should accelerate an already survivable business, so founders should first define a default path to self-sustaining operations, then model how abundant or scarce capital would change that path over the next 6 to 18 months.
FPV Ventures 合伙人 Nikunj Kothari 提醒创始人，不要假设下一轮融资必然发生。资本应当用于加速一个本就能生存的业务，因此应先建立通向自我造血的默认路径，再推演未来 6 至 18 个月资本充裕或紧缩会如何改变这条路径。
https://x.com/nikunj/status/2100008917980102863
Every CEO Dan Shipper
Dan Shipper says Every tested a foundation model that outputs probabilities rather than words and found it useful as a judge for tasks that would otherwise require a much larger model. In Every's tests it was 25 times faster and 600 times cheaper, suggesting specialized probabilistic models may become indispensable for high-volume evaluation and routing.
Every CEO Dan Shipper 表示，Every 测试了一种输出概率而非文本的 foundation model，可作为原本需要更大模型处理的 judge。在其测试中，该模型速度快 25 倍、成本低 600 倍，说明专用概率模型可能成为高频评估与路由任务的重要基础设施。
https://x.com/danshipper/status/2099947471518474522
SPC General Partner Aditya Agarwal
Aditya Agarwal traced AI marketing platform Profound back to a founder partnership formed at SPC before the idea itself existed. Profound is now valued at $1.8 billion, serves one third of the Fortune 100, and raised a $180 million Series D co-led by Sequoia and Kleiner Perkins, reinforcing the value of backing complementary founder teams before product certainty emerges.
SPC General Partner Aditya Agarwal 回顾了 AI 营销平台 Profound 的起点：在明确产品想法出现之前，两位能力互补的创始人先在 SPC 建立了合作关系。Profound 如今估值 18 亿美元，服务三分之一的 Fortune 100，并完成由 Sequoia 与 Kleiner Perkins 共同领投的 1.8 亿美元 Series D，印证了在产品确定性形成前押注互补创始团队的价值。
https://x.com/adityaag/status/2099939685657141257
Anthropic's Claude
Anthropic launched Salesforce in Claude in beta with 37 pre-built sales skills. Users can bring accounts, opportunities, and pipeline data into Claude to prepare calls, review deals, build pipeline dashboards, and send forecasts without leaving the conversation.
Anthropic 推出 Salesforce in Claude beta，内置 37 项销售 skills。用户可把客户、商机与 pipeline 数据带入 Claude，在对话中完成通话准备、deal review、pipeline dashboard 创建和 forecast 发送。
https://x.com/claudeai/status/2099876514330206578
PODCASTS
Training Data: Box's Aaron Levie: On Reinventing Yourself in the AI Age and Enterprise Diffusion
The Takeaway: The trillion-dollar AI opportunity is not merely smarter models, but the difficult work of fitting intelligence into real enterprise workflows.
Box co-founder and CEO Aaron Levie argues that the application layer remains defensible because enterprises need far more than raw intelligence: data connections, permissions, domain context, human checkpoints, process redesign, governance, and change management. Coding agents spread quickly because code is centralized, outcomes are testable, users are technical, and work happens almost entirely at a computer. Legal, sales, life sciences, and other knowledge work lack those advantages, so diffusion will be slower and create room for deeply specialized vendors.
Levie's sharpest strategic point is that AI makes building faster while simultaneously making distribution more decisive. Two founders can now create what once required a 40-person team, but every promising idea attracts competitors almost immediately. As he puts it, “There is quite literally trillion to trillions up for grab at the applied layer,” and the winners will be those that can reach enterprises, implement the technology, and sustain trust.
At Box, reinvention depends on strong data hygiene, targeting high-leverage workflows, internal AI specialists, and studying power users rather than merely tracking token consumption. The broader lesson is practical: becoming AI-first is an ongoing operating-system change, not a one-time tool rollout.
核心结论： 万亿美元级的 AI 机会不只来自更聪明的模型，更来自把智能真正嵌入企业工作流的艰难工程。
Box 联合创始人兼 CEO Aaron Levie 认为，application layer 仍具备强大护城河，因为企业需要的远不止原始智能，还包括数据连接、权限、领域上下文、人工检查点、流程重构、治理与 change management。Coding agent 扩散快，是因为代码集中、结果可测试、用户技术能力强，而且工作几乎完全发生在电脑上；法律、销售、生命科学等知识工作并不具备这些条件，因此落地会更慢，也会为深度垂直厂商留下空间。
Levie 最关键的战略判断是：AI 一方面让产品构建更快，另一方面让分发能力更加决定胜负。如今两个人就能完成过去需要 40 人的项目，但任何好点子也会立刻吸引竞争者。正如他说的：“applied layer 确实有数万亿美元的机会等待争夺。”最终赢家将是那些能进入企业、完成实施并持续建立信任的公司。
Box 的自我革新依赖良好的 data hygiene、优先改造高杠杆工作流、配置内部 AI 专家，以及研究高阶用户的真实用法，而非只看 token 消耗。更普遍的启示是：AI-first 是持续的企业 operating system 变革，不是一次工具采购。
https://www.youtube.com/playlist?list=PLOhHNjZItNnMm5tdW61JpnyxeYH5NDDx8
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders