AI Builders Digest — 2026-07-20

X / TWITTER

Swyx

Swyx pointed to Europe as an underrated AI talent pool, arguing that the region has some of the world's strongest AI engineers if you know how to surface them. He also said AEO is becoming a meaningful revenue channel for him, projecting it could drive $1M in revenue next year at the current pace.

Links: https://x.com/swyx/status/2078628617987518855, https://x.com/swyx/status/2078581967768166591

Swyx 认为欧洲 AI 人才被低估了，真正会挖掘的话，那里有一批全球顶尖 AI 工程师。他还提到 AEO 正在变成实质收入渠道，按当前速度，明年可能给他带来 100 万美元收入。

链接: https://x.com/swyx/status/2078628617987518855, https://x.com/swyx/status/2078581967768166591

OpenAI Codex & ChatGPT engineer Thibault Sottiaux

Thibault Sottiaux showed how he is using ChatGPT Work as a delegation layer: dictating a complex workflow that scans thousands of X DMs, extracts ChatGPT Work beta applicants, structures them into a spreadsheet, builds a taxonomy of use cases, and rates workflow sophistication for cohort selection. He also framed ChatGPT Work as a broad work OS for creating and hosting sites, managing email, summarizing document piles, and producing docs, sheets, and slides across Plus, Pro, Business, and Enterprise plans.

Links: https://x.com/thsottiaux/status/2078702412085498087, https://x.com/thsottiaux/status/2078697741455356367, https://x.com/thsottiaux/status/2078697631019303273

Thibault Sottiaux 展示了 ChatGPT Work 作为“委托层”的用法：直接口述一个复杂任务，让它扫描数千条 X DM，提取 ChatGPT Work beta 申请者，整理成表格，自动归类使用场景，并给工作流成熟度打分，用于挑选测试 cohort。他也把 ChatGPT Work 定位成更宽的工作操作系统：建站和托管、管理邮件、总结大量文档、生成 docs / sheets / slides，并覆盖 Plus、Pro、Business、Enterprise。

链接: https://x.com/thsottiaux/status/2078702412085498087, https://x.com/thsottiaux/status/2078697741455356367, https://x.com/thsottiaux/status/2078697631019303273

Peter Yang

Peter Yang shared a small but telling consumer AI build: he and his eight-year-old made a ChatGPT Site for multiplication practice, using ChatGPT Images for the UI and characters, adding music, and even building a timed boss level. The signal is not the app itself but the compression of kid-friendly product design, asset generation, game mechanics, and hosting into a parent-child weekend workflow.

Link: https://x.com/petergyang/status/2078638568784994686

Peter Yang 分享了一个很小但很有代表性的消费级 AI 作品：他和 8 岁女儿一起做了一个练乘法表的 ChatGPT Site，用 ChatGPT Images 生成 UI 和角色，加了音乐，还做了限时 boss 关卡。重点不是这个应用本身，而是儿童友好产品设计、资产生成、游戏机制和托管被压缩进一次亲子创作流程。

链接: https://x.com/petergyang/status/2078638568784994686

Anthropic Claude Code engineer Thariq

Thariq credited Anthropic's team for getting Fable shipped under intense time pressure, saying many people worked around the clock and that success was not guaranteed. This is a useful operating signal: Claude Code / Fable work is not just model capability, but a product and infrastructure execution push inside Anthropic.

Link: https://x.com/trq212/status/2078514180051906864

Thariq 把 Fable 的交付归功于 Anthropic 团队的高强度推进，提到很多人几乎昼夜工作，而且一开始并不确定能按时完成。这是一个有价值的执行信号：Claude Code / Fable 不只是模型能力展示，也是 Anthropic 内部产品和基础设施团队的一次硬仗。

链接: https://x.com/trq212/status/2078514180051906864

Vercel CEO Guillermo Rauch

Guillermo Rauch shared internal cybersecurity eval impressions: Kimi K3 looks top-tier despite benchmark-overfitting chatter, Sol appears meaningfully ahead at a higher cost, while Fable refused the tested tasks. His conclusion was that frontier open-weight cybersecurity capability has arrived, with defensive hardening as the intended use case. In a separate post, he argued that “AGI” is the wrong frame: AI may be better than humans at many economically relevant tasks, but human identity, judgment, care, and original voice remain the core advantage.

Links: https://x.com/rauchg/status/2078647648307880209, https://x.com/rauchg/status/2078548458714406959

Guillermo Rauch 分享了内部 cyber eval 的观察：尽管外界怀疑 Kimi K3 存在 benchmark overfit，但它在隐蔽评测里表现为顶级；Sol 成本更高但能力明显领先；Fable 则几乎拒绝完成测试任务。他的结论是，前沿 open-weight 模型的 cybersecurity 能力已经到来，适合用于防御性加固。另一条里，他认为 “AGI” 这个词已经不准了：AI 在很多经济任务上会强于人，但人的身份、判断、关怀和独特表达仍然是核心优势。

链接: https://x.com/rauchg/status/2078647648307880209, https://x.com/rauchg/status/2078548458714406959

Box CEO Aaron Levie

Aaron Levie argued that AI value will not accrue only to a few frontier labs. He sees a healthy ecosystem forming around model customization and inference, applied AI workflow companies, vertical labs in domains like life sciences and finance, infrastructure for agent orchestration and governance, and services firms that handle enterprise change management. He also pushed back on model gatekeeping, saying competition with China requires faster progress, broader diffusion, infrastructure buildout, and stronger US open source rather than slowing the ecosystem down.

Links: https://x.com/levie/status/2078567715544121815, https://x.com/levie/status/2078481578779685245

Aaron Levie 认为 AI 价值不会只流向少数 frontier labs。他看到一个健康生态正在形成：模型定制与推理、应用层 workflow 公司、生命科学和金融等垂直实验室、agent 编排与治理基础设施、以及帮助企业完成 change management 的新服务公司。他也反对单纯封锁模型，认为面对中国竞争，美国需要更快进步、更广泛扩散、更多基础设施建设和更强的 US open source，而不是拖慢整个生态。

链接: https://x.com/levie/status/2078567715544121815, https://x.com/levie/status/2078481578779685245

FirstMark Capital VC Matt Turck

Matt Turck pushed back on the recurring claim that the model layer is commoditizing, joking that people have said the same thing in 2024, 2025, and 2026 while the model layer remains very much not commoditized. The underlying point: frontier model differentiation, distribution, infra cost, and ecosystem control still matter more than many application-layer narratives admit.

Link: https://x.com/mattturck/status/2078520552680046920

Matt Turck 反驳了“模型层正在商品化”的反复叙事，调侃 2024、2025、2026 大家都这么说，但模型层仍然没有真正 commoditize。背后的判断是：frontier model 的差异化、分发、基础设施成本和生态控制权，仍然比很多应用层叙事承认的更重要。

链接: https://x.com/mattturck/status/2078520552680046920

Builder Zara Zhang

Zara Zhang recommended that everyone build a personal eval set for AI models: a few tasks that match their own daily work and life, rather than relying only on industry benchmarks. She also named the main enterprise adoption blocker clearly: people who understand AI often do not understand the business, while people who understand the business often do not understand AI.

Links: https://x.com/zarazhangrui/status/2078666187026911488, https://x.com/zarazhangrui/status/2078492577788268549

Zara Zhang 建议每个人都建立自己的 AI model personal eval set：挑几个真正贴近日常工作和生活的任务，而不是只看行业 benchmark。她也直接指出企业 AI 落地的最大障碍：懂 AI 的人不懂业务，懂业务的人不懂 AI。

链接: https://x.com/zarazhangrui/status/2078666187026911488, https://x.com/zarazhangrui/status/2078492577788268549

PODCASTS

Unsupervised Learning: Ep 90: AI Pioneer Jürgen Schmidhuber on the State of AI Today

The Takeaway: Jürgen Schmidhuber’s core argument is that today’s screen-bound LLMs are impressive but incomplete, because true general intelligence needs agents that act in the physical world, invent their own experiments, and generate their own training data through curiosity.

Schmidhuber, one of the long-running pioneers behind modern neural networks and meta-learning ideas, is optimistic about AI technology but skeptical of the market structure around today’s model companies. He argues that current LLMs are deeply human-biased because they are trained on web data selected by human interest. The next important step is closer to an “artificial scientist”: a system that learns by acting, predicting consequences, finding compressible patterns near the edge of what it already understands, and improving its own world model. He also pushes against the idea that recursive self-improvement automatically becomes a company moat, because many current approaches are scaled-down versions of older self-modifying systems and remain constrained by gradient-based optimization.

Link: https://www.youtube.com/watch?v=RKjR8DQ40po

核心判断：Jürgen Schmidhuber 的重点是，今天这些“屏幕里的 LLM”已经很强，但仍不完整。真正的 general intelligence 需要能在物理世界行动、自己设计实验、通过 curiosity 生成训练数据的 agent。

Schmidhuber 是现代神经网络和 meta-learning 的长期先驱之一。他对 AI 技术本身乐观，但对今天模型公司的市场结构保持怀疑。他认为当前 LLM 高度“人类偏置”，因为它们主要训练在由人类兴趣筛选过的 web 数据上。下一步更像“人工科学家”：系统通过行动、预测后果、寻找自己刚好能学会的可压缩模式，并持续改进 world model。他也反对把 recursive self-improvement 自动等同于公司护城河，因为许多当前方法只是早期 self-modifying systems 的缩小实用版，仍受 gradient-based optimization 的约束。

链接: https://www.youtube.com/watch?v=RKjR8DQ40po

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