AI Builders Digest | 2026-09-24

X / TWITTER

Boris Cherny | Claude Code, Anthropic

Boris Cherny used Opus 5.5 with Lean to formally verify the Claude Agent SDK. A few short prompts produced 16 PRs that fixed bugs and race conditions; he also combines Lean with TLA+ to inspect data flow, concurrency, and state management. In a separate benchmark, Opus 5.5 ported HAProxy from C to Rust in 9.5 hours, versus 12 hours for Fable 5.1, at 51% lower cost, with both passing nearly all tests.

Boris Cherny 使用 Opus 5.5 和 Lean 对 Claude Agent SDK 进行形式化验证，几个简短 prompt 就产出了 16 个 PR，修复了多处 bug 与竞态条件。他还会结合 Lean 与 TLA+ 检查数据流、并发和状态管理。另一项测试中，Opus 5.5 用 9.5 小时将 HAProxy 从 C 移植到 Rust，Fable 5.1 用时 12 小时，前者成本低 51%，两者都通过了几乎全部测试。

Sources: formal verification · HAProxy benchmark

Thibault Sottiaux | Codex & ChatGPT, OpenAI

Thibault Sottiaux announced GPT-6 Sol and Luna, highlighting broad capability and writing improvements. OpenAI also permanently cut their API prices by 50%, expanding the range of economically viable use cases, and added a banked reset for Plus, Pro, and Business users.

Thibault Sottiaux 宣布推出 GPT-6 Sol 和 Luna，强调其综合能力与写作质量均有明显提升。OpenAI 同时将两款模型的 API 价格永久下调 50%，让更多场景具备经济可行性，并为 Plus、Pro 和 Business 用户增加一次可储备的额度重置。

Source: launch details

Cat Wu | Claude Code & Cowork, Anthropic

Cat Wu said Claude Opus 5.5 is now the default model in Claude Code and the Claude app, including Cowork, for Pro, Max, and Team plans. Products default to medium effort, which she describes as comparable to Fable 5.1 in intelligence but faster; rate limits should go 25% further than with Opus 5.

Cat Wu 表示，Claude Opus 5.5 已成为 Claude Code 和 Claude 应用（包括 Cowork）在 Pro、Max 与 Team 套餐中的默认模型。产品默认采用 medium effort，她称其智能水平与 Fable 5.1 相当但速度更快，且相较 Opus 5，用户的 rate limit 可多支撑约 25% 的使用量。

Source: default-model announcement

Thariq | Claude Code, Anthropic

Thariq argues that stronger models should not be used to push ten times more features into production. Their real value is giving teams more room to understand users, run experiments, build prototypes, and learn unfamiliar domains before shipping what actually works. For game builders, 3D generation can accelerate visualization, but it cannot replace a satisfying core game loop.

Thariq 认为，更强的模型不应该被用来向生产环境堆入十倍功能。它真正的价值，是让团队有更多空间理解用户、开展实验、构建原型，并在发布前补足陌生领域的认知。对游戏开发者而言，3D 生成能加速想象落地，但无法替代真正令人满意的核心玩法循环。

Sources: product-building principle · game-loop reminder

Guillermo Rauch | CEO, Vercel

Guillermo Rauch's fresh Next.js eval put Opus 5.5, GPT-6 Sol, and Fable 5.1 at 97%, with Grok 4.7 close behind at 94% while costing 2 to 7 times less. He also argues that generative software changes product longevity: if a beloved service disappears, users can increasingly generate and deploy their own replacement.

Vercel CEO Guillermo Rauch 的最新 Next.js eval 显示，Opus 5.5、GPT-6 Sol 与 Fable 5.1 均达到 97%，Grok 4.7 以 94% 紧随其后，但成本低 2 到 7 倍。他还提出，生成式软件正在改变产品寿命：当一个喜爱的服务消失时，用户将越来越能够自行生成并部署替代品。

Sources: Next.js eval · software permanence

Alex Albert | Research, Anthropic

Alex Albert demonstrated Opus 5.5's 3D and vision capabilities with a historically sourced reconstruction of San Francisco's Market Street before the 1906 earthquake. The prompt required a structured source file, confidence levels for facts, reusable Blender Python generators, no downloaded assets, and a final ten-second street video, showing how ambitious outputs improve when research provenance is part of the specification.

Anthropic 研究员 Alex Albert 用一个基于历史资料的项目展示了 Opus 5.5 的 3D 与视觉能力：重建 1906 年大地震前的旧金山 Market Street。Prompt 要求先建立结构化来源文件，为每项事实标注可信度，用 Blender Python 编写可复用生成器，禁止下载外部素材，并最终输出十秒街景视频。这说明，把研究溯源直接写进任务规格，能显著提升复杂生成结果的可靠性。

Sources: full prompt · result overview

Aaron Levie | CEO, Box

Aaron Levie frames falling model prices as Jevons paradox for agents: lower cost per task expands demand and unlocks broad deployment across security, logs, data processing, and workflow swarms. Box's Opus 5.5 tests reported 63% fewer tokens, 42% less verbosity, and 30% faster execution than Opus 5, with task-accuracy gains ranging from 15% to 65% across clinical, financial, account, and cloud-cost analysis.

Box CEO Aaron Levie 将模型降价视为 agent 领域的 Jevons paradox：单任务成本越低，需求反而越大，并推动 agent 在安全、日志、数据处理和工作流集群中广泛部署。Box 对 Opus 5.5 的测试显示，相比 Opus 5，其 token 用量减少 63%、冗长度降低 42%、执行速度提升 30%；在临床、金融、客户与云成本分析任务中，准确率提升 15% 到 65%。

Sources: agent economics · Box enterprise tests

Garry Tan | President & CEO, Y Combinator

Garry Tan says Capy lets him ship PRs much faster than using Codex or Claude Code alone. His broader point is that the next adoption challenge is teaching people how to prompt and use AI deeply enough to expand what they can pursue, then channeling that leverage toward more meaningful problems.

Y Combinator 总裁兼 CEO Garry Tan 表示，Capy 让他提交 PR 的速度明显快于单独使用 Codex 或 Claude Code。他更广泛的判断是，下一阶段的普及挑战在于教会人们真正掌握 prompt 并深度使用 AI，从而扩展个人能力边界，再把这种杠杆用于解决更多有意义的问题。

Sources: Capy workflow · AI adoption

Nikunj Kothari | Partner, FPV Ventures

Nikunj Kothari warns that large funding headlines increasingly conceal SPV participation, tranched valuations, and ambiguous revenue definitions. His takeaway for founders and investors is simple: headline round size is no longer a reliable proxy for financing quality or company fundamentals.

FPV Ventures 合伙人 Nikunj Kothari 警告，大额融资新闻越来越多地掩盖 SPV 参与、分阶段估值以及模糊的收入定义。对创始人和投资者而言，核心结论是：新闻标题中的融资规模，已不能可靠代表融资质量或公司的真实基本面。

Source: funding-market caution

Peter Steinberger | OpenClaw & OpenAI

Peter Steinberger reported that after ChatGPT began crashing on macOS 27, Astra traced the problem to an approximately 14-year-old bug in libuv. It is a compact example of AI-assisted debugging finding faults buried deep in mature infrastructure rather than merely generating application code.

OpenClaw 与 OpenAI 的 Peter Steinberger 表示，在 ChatGPT 升级到 macOS 27 后偶发崩溃时，Astra 最终定位到 libuv 中一个约 14 年前的 bug。这个案例说明，AI 辅助调试的价值不仅是生成应用代码，还能深入成熟基础设施，发现长期潜伏的问题。

Source: libuv debugging case

Aditya Agarwal | General Partner, SPC

Aditya Agarwal points to Waymo's extensive eval and testing infrastructure as the foundation for safely deploying two-ton autonomous machines on public roads. He sees autonomous driving as the original AI safety debate and a useful reminder that confidence comes from rigorous systems-level validation, not model capability alone.

SPC General Partner Aditya Agarwal 指出，Waymo 大规模的 eval 与测试基础设施，是其能在公共道路安全部署两吨重自动驾驶机器的根基。他把自动驾驶视为最早的 AI safety 讨论，并提醒我们：可信部署来自严谨的系统级验证，而不只是模型能力本身。

Source: Waymo safety infrastructure

PODCASTS

No new podcast episodes in this run.

本次运行没有新的 podcast 节目。

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