AI Builders Digest — 2026-10-15

X / TWITTER

Swyx (AI Engineer / Latent Space)

Swyx tested the “death of SaaS” against a real target: a global hackathon challenged teams to clone a $43,000-per-year SaaS product in one weekend using any model or agent harness. After 72 submissions, human and AI evaluations, and roughly 100 billion tokens, the results expose what remains difficult about replacing nontrivial mid-market software rather than merely generating a convincing demo.

Swyx 用真实项目检验了“SaaS 已死”的说法：一场全球黑客松要求参赛者在一个周末内，使用任意模型或 agent harness 克隆一款年费 43,000 美元的 SaaS。72 份作品、人工与 AI 双重评测以及约 1000 亿 token 的消耗表明，替代有实际业务复杂度的中型 SaaS，远比做出一个看似可用的 demo 更难。

https://x.com/swyx/status/2110065576764686421

Josh Woodward, VP at Google Labs / Gemini

Josh Woodward showcased Gemini 4 Argon turning a 24-word prompt into a custom synthesizer controlled by video of horses, water, and hand movements. He also highlighted its ability to move from generative design to a manufacturable 96-millimeter furniture component, positioning the model as a tool for rapidly translating compact intent into interactive software and physical designs.

Google Labs / Gemini VP Josh Woodward 展示了 Gemini 4 Argon 如何用一条仅 24 个词的 prompt，生成由奔马、水波和手势视频控制的定制合成器。他还展示了模型从生成式设计走向可制造的 96 毫米家具部件，体现出模型把简短意图快速转化为交互软件与实体设计的能力。

https://x.com/joshwoodward/status/2110179883045560674

https://x.com/joshwoodward/status/2110089824573272547

Thibault Sottiaux, Codex & ChatGPT at OpenAI

Thibault Sottiaux said a desktop-app storage optimization for ChatGPT and Codex should reduce on-disk state by 60%. It is a small infrastructure release, but one driven directly by reports that heavy users were running out of disk space, a useful reminder that local resource management becomes a product feature as agent usage scales.

OpenAI 的 Thibault Sottiaux 表示，ChatGPT 与 Codex 桌面端已优化本地状态存储，预计磁盘占用可减少 60%。这看似是小型基础设施更新，却直接回应了重度用户磁盘被占满的问题，也说明随着 agent 使用规模扩大，本地资源管理本身正在成为产品能力。

https://x.com/thsottiaux/status/2110151780478927344

Amjad Masad, Replit CEO

Amjad Masad argues that even if frontier research stopped today, AI would still need another decade to diffuse through the economy. Replit is therefore working with solution partners to compress the adoption gap, suggesting that implementation and distribution, not only model capability, are now major constraints on realized AI value.

Replit CEO Amjad Masad 判断，即使前沿研究今天完全停止，AI 仍需要约十年才能充分渗透进经济体系。因此 Replit 正通过 solution partners 缩短采用周期，这意味着 AI 价值落地的主要瓶颈，正从模型能力转向实施与分发。

https://x.com/amasad/status/2110128937049452575

Guillermo Rauch, Vercel CEO

Guillermo Rauch calls coding agents AI’s killer application and says Google’s data, engineering talent, and compute make it dangerous to underestimate. His practical recommendation is model agnosticism: rapid back-to-back model releases make a gateway or flexible model-routing layer strategically more valuable than dependence on a single provider.

Vercel CEO Guillermo Rauch 将 coding agent 称为 AI 的 killer app，并认为 Google 拥有数据、工程人才与算力优势，不应被低估。他给出的实操建议是保持 model agnostic：当强模型密集发布时，gateway 或灵活的模型路由层，比绑定单一供应商更具战略价值。

https://x.com/rauchg/status/2110044152482365747

Aaron Levie, Box CEO

Aaron Levie sees cybersecurity as one of the largest future markets for AI agents. Agents expand the attack surface, but defensive swarms can also review code, test exploit permutations, and triage incidents at a scale human teams cannot match, creating opportunities across models, data, networks, identity, endpoints, SDLC, and enterprise data protection.

Box CEO Aaron Levie 认为，网络安全将成为 AI agent 最大的未来市场之一。Agent 会扩大攻击面，但防御型 agent swarm 也能以人类团队无法达到的规模审查代码、测试漏洞组合并分诊安全事件，从模型、数据、网络、身份、终端、SDLC 到企业数据保护都会出现新机会。

https://x.com/levie/status/2110227571090788526

Box’s internal Complex Work evaluation found Gemini 4 Argon improved 14 percentage points over Gemini 3.1 Pro across enterprise-document tasks. The largest cited gains came in technology (+37), consumer products (+33), healthcare (+25), financial services (+18), and energy (+16); Levie emphasizes that Argon not only reasons better but more consistently carries tasks through to a result matching its own diagnosis.

Box 的内部 Complex Work 评测显示，Gemini 4 Argon 在企业文档任务上较 Gemini 3.1 Pro 整体提升 14 个百分点。公开的分行业增幅包括科技 +37、消费品 +33、医疗 +25、金融服务 +18、能源 +16；Levie 强调，Argon 不只是推理更强，也更能把任务完整执行到与自身诊断一致的结果。

https://x.com/levie/status/2110046686198550539

Garry Tan, Y Combinator President & CEO

Garry Tan recommends giving agent harnesses access to a real browser with the user’s real credentials, calling it a major unlock and pointing to AsideAI. The value is obvious for end-to-end execution, though the same setup raises the stakes for credential isolation, permissions, audit logs, and reversible actions.

Y Combinator President & CEO Garry Tan 推荐让 agent harness 使用带真实用户凭证的真实浏览器，并将 AsideAI 视为一个重要突破。这能显著增强端到端执行能力，但也同步提高了凭证隔离、权限控制、审计日志与可逆操作的重要性。

https://x.com/garrytan/status/2110053836945756234

Nikunj Kothari, FPV Ventures Partner

Nikunj Kothari points to Railway as an “n of 1” company shaped by founder Jake Cooper’s unconventional approach to scaling, selling, hiring, and building. The broader builder lesson is that exceptional companies often come from a coherent operating system tailored to the founder, not from copying standard startup playbooks.

FPV Ventures Partner Nikunj Kothari 将 Railway 视为一家由创始人 Jake Cooper 独特的扩张、销售、招聘与产品方法塑造的“n of 1”公司。更普遍的启示是，卓越公司往往源于与创始人高度匹配且内部一致的 operating system，而不是照搬标准创业手册。

https://x.com/nikunj/status/2110137649168281871

Dan Shipper, Every CEO

Dan Shipper’s testing offers a sharp counterpoint to broad Gemini 4 enthusiasm: Gemini 4 Argon scored only 21% on Every’s senior-engineer benchmark, below Sol 6 and Muse. Together with Box’s strong enterprise results, this suggests model quality is highly workload-dependent and reinforces the need for domain-specific evaluations before changing production defaults.

Every CEO Dan Shipper 的测试为普遍看好 Gemini 4 的声音提供了鲜明反例：Gemini 4 Argon 在 Every 的 senior engineer benchmark 上仅得 21%，低于 Sol 6 和 Muse。结合 Box 在企业任务中的优秀结果，更合理的结论是模型表现高度依赖工作负载，生产环境切换默认模型前必须进行领域专属评测。

https://x.com/danshipper/status/2110056016872427944

PODCASTS

Training Data: David Silver and Lasse Espeholt of Ineffable Intelligence: A Superlearner for the Age of Experience

The Takeaway: David Silver and Lasse Espeholt are betting that superintelligence will emerge from agents that learn autonomously through experience, not from extending today’s human-data and language-centered LLM pipeline.

Silver, the reinforcement-learning researcher behind AlphaGo, left DeepMind to co-found Ineffable Intelligence because he did not want to become the “100,001st person” pursuing the same pretraining and post-training path. Their proposed superlearner must have both depth, progressing from no knowledge to discoveries beyond humanity, and breadth, learning from simulated, digital, or physical experience. “Everything we learned, we discovered it for ourselves,” Silver says; AI has not yet reproduced that open-ended process.

The lab is building a large-scale RL stack and “crucibles,” simulated worlds whose physics and multi-agent interactions generate complexity rather than presenting hand-authored task lists. Agents receive relatively simple objectives and must discover tools, communication, science, or mathematics without breadcrumb rewards. The deliberate rejection of human data sacrifices short-term performance to avoid importing human priors and ceilings.

Silver argues the timing is now plausible because compute, scaling knowledge, and AI-assisted research have converged; Ineffable expects experiments roughly 100,000 times larger than early emergent-agent work such as OpenAI’s hide-and-seek. Simulation also provides a faster and safer place to study alignment, including non-language monitoring and the conditions under which powerful agents coexist peacefully. The wager is ambitious but concrete: build environments rich enough for discovery, then measure whether increasingly general intelligence emerges without humans specifying the path.

核心结论： David Silver 与 Lasse Espeholt 押注，超级智能将来自通过经验自主学习的 agent，而不是继续扩展今天依赖人类数据、以语言为中心的 LLM 路线。

AlphaGo 背后的强化学习研究者 Silver 离开 DeepMind、共同创办 Ineffable Intelligence，是因为他不想成为沿 pretraining 与 post-training 路线前进的“第 100,001 个人”。他们定义的 superlearner 必须同时具备深度与广度：能从零知识一路发现超越人类的知识，也能从模拟、数字或物理环境中学习。Silver 说：“Everything we learned, we discovered it for ourselves.” 而 AI 尚未复现这种开放式发现过程。

团队正在构建大规模 RL 技术栈以及名为“crucibles”的模拟世界。复杂性来自物理系统和多 agent 互动，而非人工编写的任务清单。Agent 只获得相对简单的目标，需要在没有面包屑式 reward 的情况下自行发现工具、通信、科学或数学。主动放弃人类数据会牺牲短期性能，但目的是避开人类先验与能力上限。

Silver 判断现在时机已成熟，因为算力、scaling 规律以及 AI 辅助研究正在汇合。Ineffable 预计可运行规模约为 OpenAI hide-and-seek 等早期涌现式 agent 工作 10 万倍的实验。模拟环境也能更快、更安全地研究 alignment，包括非语言监控，以及强大 agent 和平共存所需的条件。这是一次雄心巨大但路径具体的押注：先构建足够丰富、可产生发现的环境，再检验无需人类规定路径的通用智能能否持续涌现。

https://www.youtube.com/playlist?list=PLOhHNjZItNnMm5tdW61JpnyxeYH5NDDx8

Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders