AI 建造者日报 — 2026年7月2日
📌 X/TWITTER
Claude (Anthropic’s AI Assistant)
Sonnet 5 is now the default on Free and Pro plans, available to Max, Team, and Enterprise users. It’s a substantial improvement over Sonnet 4.6 on reasoning, tool use, coding, and knowledge work, with performance close to Opus 4.8 at lower prices. Early access partners found Sonnet 5 finishes complex tasks where previous Sonnets stopped short, checks its own output without being asked, and does all its agentic work at an attractive price point.
Claude（Anthropic AI 助手）：Sonnet 5 现已成为 Free 和 Pro 计划的默认模型，Max、Team 和 Enterprise 用户也可使用。相比 Sonnet 4.6，在推理、工具使用、编码和知识工作方面有显著提升，性能接近 Opus 4.8 但价格更低。早期合作伙伴发现 Sonnet 5 能完成之前 Sonnet 模型半途而止的复杂任务，无需提示即可自我检查输出，并以有竞争力的价格完成所有 agentic 工作。
Box CEO Aaron Levie
Fable’s release and the upcoming GPT-5.6 are setting the precedent for what frontier model releases with significant coding and cyber capabilities will look like going forward. On Sonnet 5: Box ran it through their Complex Work Eval benchmark and found it holds frontier-class quality on complex multi-step enterprise document work, pulling ahead of Sonnet 4.6 in several core enterprise domains. On AI and jobs: contrary to expectations, Ramp found more AI adoption correlates with more headcount growth. Box’s own survey of 1,600+ companies found 58% expect AI to increase hiring.
Box CEO Aaron Levie：Fable 的发布和即将到来的 GPT-5.6 正在为具备强大编码和网络能力的前沿模型发布树立先例。关于 Sonnet 5：Box 通过其 Complex Work Eval 基准测试发现，Sonnet 5 在复杂多步骤企业文档工作中保持前沿级质量，在多个核心企业领域超越 Sonnet 4.6。关于 AI 与就业：与预期相反，Ramp 发现 AI 采用率越高的公司，人员增长越快。Box 对 1600+ 家企业的调查显示，58% 的受访者预期 AI 将增加招聘。
- https://x.com/levie/status/2072172275017879829
- https://x.com/levie/status/2072046374045249671
- https://x.com/levie/status/2071992799109824562
Vercel CEO Guillermo Rauch
Vercel Services launched: you can now collocate a Python backend API, an ExpressJS server, and a React SPA in one Vercel project — run locally with vc dev, deploy and rollback together, with unified observability and internal networking. Also announced partnership with Shopify to push the agentic web forward.
Vercel CEO Guillermo Rauch：Vercel Services 发布——现在可以在一个 Vercel 项目中统一管理 Python 后端 API、ExpressJS 服务器和 React SPA，用 vc dev 本地运行，统一部署和回滚，支持统一可观测性和内部网络。同时宣布与 Shopify 合作推进 agentic web。
Anthropic Claude Code Team — Boris Cherny
Claude Desktop is now available on Linux.
Anthropic Claude Code 团队成员 Boris Cherny：Claude Desktop 现已支持 Linux。
Replit CEO Amjad Masad
Most AI workloads today run on generic hardware designed pre-LLMs. Etched is the first system designed from the ground up for modern inference — highlighting the growing shift toward purpose-built AI hardware.
Replit CEO Amjad Masad：当前大多数 AI 工作负载运行在 LLM 时代之前设计的通用硬件上。Etched 是首个从零为现代推理设计的系统——凸显了向专用 AI 硬件转变的趋势。
South Park Commons GP Aditya Agarwal
“It is a very strange state of the world where the models powering innovation in the USA are Chinese open source models.” A striking observation on the shifting geopolitics of AI.
South Park Commons GP Aditya Agarwal：“这是一个非常奇怪的世界状态——推动美国创新的模型是中国的开源模型。“对 AI 地缘政治格局变化的犀利观察。
Linear Head of Product Nan Yu
On “distillation”: if we take OpenAI’s definition seriously, “the entirety of Cursor’s training data was distilled from Claude in the early days” — a provocative reframing of what distillation really means in the AI tooling ecosystem.
Linear 产品负责人 Nan Yu：关于"蒸馏"的定义——如果以 OpenAI 的逻辑来审视，“Cursor 在早期的全部训练数据都是从 Claude 蒸馏出来的”——对 AI 工具生态中"蒸馏"含义的挑衅性重新定义。
Former Google Product Leader Madhu Guru (Gemini/Veo)
The biggest challenge for traditional PMs adapting to AI-native building is a lack of “magical thinking.” A decade of frameworks, agile, and metric obsession has led to constraint-first, incremental thinking. He used to combat this by asking teams to imagine technology from 100 years in the future.
前 Google 产品负责人 Madhu Guru（Gemini/Veo）：传统 PM 适应 AI 原生开发的最大挑战是缺乏"魔法思维”。十年的框架、敏捷和指标痴迷导致了约束优先的渐进式思维。他过去通过让团队想象 100 年后的技术来对抗这种倾向。
Y Combinator CEO Garry Tan
GBrain is most useful at 10,000+ markdown files in your personal or company knowledge base — reinforcing the value of AI tools at scale, not just for small projects.
Y Combinator CEO Garry Tan：GBrain 在个人或公司知识库达到 10,000+ 个 markdown 文件时最为有用——强调 AI 工具在大规模场景下的价值，而非仅适用于小型项目。
OpenClaw/OpenAI — Peter Steinberger
Price per token ≠ cost per task. A simple but important reminder as the industry obsesses over per-token pricing while real-world task completion cost is what matters.
OpenClaw/OpenAI Peter Steinberger：每 token 价格 ≠ 每任务成本。在行业痴迷于每 token 定价时，实际任务完成成本才是真正重要的——简洁而重要的提醒。
🎙️ 播客
Training Data: Why Hardware-Software Co-Design Is AI’s Real 100x — Dylan Patel of SemiAnalysis
The Takeaway: The biggest gains in AI won’t come from scaling models alone — they’ll come from hardware-software co-design, where model architecture and chip design evolve together.
Dylan Patel, founder of SemiAnalysis (a research firm reportedly surpassing $100M in revenue), grew up in a family motel business, taught himself hardware by fixing a Red Ring of Death Xbox 360 at age 8, and became a Reddit moderator for hardware forums by 12. After working as a quant, he built SemiAnalysis into the premier semiconductor research firm.
His core argument: different AI labs are converging on fundamentally different model architectures, and those architectures are co-designed with specific hardware. OpenAI’s models are increasingly sparse, optimized for NVIDIA GPUs with NVLink switch topology. Anthropic and Google’s models are denser and better suited for TPU architectures, where Google’s ICI can connect 8,000 chips at super high bandwidth without switches. “OpenAI’s models are much more sparse and that has benefits, and then Anthropic’s are still sparse but more dense in general, and that has different benefits.”
On the CUDA moat: it’s partially disentangling because AI models are now good enough at coding to write custom kernels. The real lock-in isn’t CUDA itself — it’s that the downstream ecosystem (open-source models, inference providers, RL companies) is optimized for NVIDIA hardware because the leading open models were co-designed for it.
On Cerebras: “Really innovative” for fast inference, but if frontier models hit 10+ trillion parameters with million-token context windows, SRAM-based chips may struggle to fit them.
On Jensen Huang’s strategy: “Jensen absolutely hates a world where all the hyperscalers have all the power.” That’s why NVIDIA invests in Neo Clouds (CoreWeave, Crusoe) and Neo Labs — to create a multipolar world where no single hyperscaler can dictate terms. “A world where OpenAI, Anthropic, and Google models are the only models is one in which he’s screwed.”
The most memorable quote: “You throw a bunch of bait into the water and the best fish will figure out and survive.” — describing NVIDIA’s strategy of seeding dozens of Neo Clouds and Labs, knowing most will fail but a few will emerge as major players.
Training Data: Why Hardware-Software Co-Design Is AI’s Real 100x — SemiAnalysis Dylan Patel
核心观点：AI 最大的突破不会仅来自模型规模扩展——它将来自硬件-软件协同设计，模型架构和芯片设计共同进化。
Dylan Patel，SemiAnalysis 创始人（据报道营收已超 1 亿美元），在家庭汽车旅馆长大，8 岁时通过修理一台"红环死机"的 Xbox 360 自学硬件，12 岁成为 Reddit 硬件论坛版主。在做过两年量化分析师后，他将 SemiAnalysis 打造成了首屈一指的半导体研究公司。
他的核心论点：不同 AI 实验室正在趋近根本不同的模型架构，而这些架构与特定硬件协同设计。OpenAI 的模型越来越稀疏化，针对 NVIDIA GPU 的 NVLink 交换拓扑优化。Anthropic 和 Google 的模型更密集，更适合 TPU 架构——Google 的 ICI 可以在无交换机的情况下以超高带宽连接 8000 个芯片。“OpenAI 的模型稀疏得多，这有它的好处；Anthropic 的模型虽也稀疏，但总体上更密集，这又有不同的好处。”
关于 CUDA 护城河：它正在部分瓦解，因为 AI 模型现在已擅长编写自定义内核。真正的锁定不是 CUDA 本身——而是下游生态（开源模型、推理服务商、RL 公司）已针对 NVIDIA 硬件优化，因为领先的开源模型是与其协同设计的。
关于 Cerebras：“非常创新"的快速推理方案，但如果前沿模型达到 10 万亿以上参数并配备百万 token 上下文窗口，基于 SRAM 的芯片可能难以承载。
关于黄仁勋的战略：“Jensen 绝对痛恨所有超大规模云厂商掌握全部权力的世界。“这就是为什么 NVIDIA 投资 Neo Clouds（CoreWeave、Crusoe）和 Neo Labs——创造一个多极世界，没有任何单一超大规模厂商可以发号施令。“一个只有 OpenAI、Anthropic 和 Google 模型的世界，他将陷入困境。”
最难忘的一句话：“你把一堆鱼饵扔进水里，最好的鱼会找出办法并存活下来。"——描述 NVIDIA 播种数十家 Neo Clouds 和 Labs 的策略，明知大部分会失败，但少数将崛起为重要玩家。
通过 Follow Builders 生成: https://github.com/zarazhangrui/follow-builders