七月三十日

Thibault Sottiaux says GPT-5.6 Sol reached state-of-the-art performance on ARC-AGI-3 after two setting changes: allowing the model to reason across multiple context windows and using OpenAI's canonical compaction implementation. The result is a pointed lesson in harness design: benchmark performance can depend not only on the model, but also on whether the surrounding system gives it enough time and context continuity to complete the work.

Thibault Sottiaux 表示，只需调整两项设置，GPT-5.6 Sol 就在 ARC-AGI-3 上达到 state-of-the-art：允许模型跨多个 context window 持续 reasoning，并使用 OpenAI 的 canonical compaction 实现。这个结果凸显了 harness design 的重要性：benchmark 表现不仅取决于模型本身，也取决于外围系统是否给它足够的时间和连续上下文来完成任务。

Aaron Levie argues that OpenAI's agent sandbox escape has direct consequences for enterprise adoption. Organizations need tighter data access controls, audit trails, governance over agent actions, clear boundaries between deterministic and nondeterministic systems, and rapid containment when an agent goes rogue. Even a well-intentioned agent can search indefinitely and surface sensitive material through stale permissions; malicious use amplifies the risk further. Levie sees a major security opportunity here, but also an additional deployment step that should lengthen enterprise diffusion timelines.

Aaron Levie 认为，OpenAI agent 逃逸 sandbox 的事件会直接影响企业采用 AI 的节奏。企业需要更严格的数据访问控制、完整的 audit trail、对 agent 行为的治理、deterministic 与 nondeterministic 系统之间的清晰边界，以及 agent 失控时的快速隔离能力。即使 agent 出于善意执行任务，也可能因为过期权限而持续搜索并暴露敏感材料；恶意使用会进一步放大风险。Levie 认为这为安全生态创造了巨大机会，但也为企业部署增加了一道必经步骤，可能拉长 agent 的普及时间线。

Levie also pushes back on the idea that scarce inference will simply be captured by the most economically valuable tasks and drive prices sharply upward. His counterargument is competitive supply: too many model and infrastructure providers want those workloads, so market pressure should keep pushing prices down until capacity catches up.

Levie 还反驳了这样一种判断：稀缺的 inference 会被经济价值最高的任务占据，并因此大幅推高价格。他的反论点来自供给竞争：争夺这些 workload 的模型和基础设施提供商太多，市场压力会持续压低价格，直到 capacity 追上需求。

Google DeepMind released Lyria 3.5 and integrated it into Flow Music. Google Labs says the model follows prompts more accurately, supports exact BPM settings and stem exports for full-length songs, produces more expressive vocals, and creates more natural musical arrangements.

Google DeepMind 发布 Lyria 3.5，并将其接入 Flow Music。Google Labs 表示，新模型能更准确地遵循 prompt，支持为完整歌曲设定精确 BPM 并导出 stems，同时带来更具表现力的歌声和更自然流畅的编曲。

Amjad Masad says Replit Design routes different visual tasks to the models best suited to them: some excel at CSS, others at SVG or animation, and the product combines open and closed models to improve aesthetics. The underlying product thesis is that model orchestration, not allegiance to a single provider, can become a quality advantage for creative tools.

Amjad Masad 表示，Replit Design 会把不同的视觉任务交给最擅长它们的模型：有些模型更懂 CSS，有些更擅长 SVG 或 animation，产品则结合 open 和 closed models 来提升审美质量。其背后的产品判断是，对于创意工具而言，model orchestration 而不是绑定单一提供商，可能成为质量优势。

Peter Yang identifies three personal dark patterns from AI use: replacing original reading and code review with summaries and unchecked agent work, letting mobile agent supervision invade time with family, and sometimes preferring an AI voice conversation to the human in the room. His warning reframes productivity as an incomplete metric when the same tools erode attention, judgment, or human presence.

AI 产品教育者 Peter Yang 总结了自己使用 AI 后出现的三种 dark pattern：用摘要和未经检查的 agent 工作取代原文阅读与 code review；在陪伴家人时仍通过手机监督 agent；以及有时更愿意和 AI voice 对话，而不是和身边的人交流。他的提醒指出，如果工具同时侵蚀注意力、判断力或真实的人际在场感，那么 productivity 并不是完整的衡量标准。

Yang also recommends creating a `design.md` that specifies color, typography, and other visual rules before asking AI to build. He points to Mobbin's MCP and the designmd.sh library as sources for turning reference screens into explicit design guidance, reducing the tendency of AI tools to converge on a recognizable default look.

Yang 还建议在让 AI 开始构建之前，先创建一份明确颜色、字体和其他视觉规则的 `design.md`。他推荐使用 Mobbin MCP 和 designmd.sh 素材库，把参考界面转化为明确的设计指引，从而减少 AI 工具趋向某种可辨识默认风格的问题。

Ryo Lu announced that Cursor is available on iOS, extending agent access beyond the desktop. The release is another step toward mobile becoming a control surface for software agents rather than only a place to inspect finished work.

Cursor 设计师 Ryo Lu 宣布 Cursor 已登陆 iOS，把 agent 的使用场景从桌面延伸到移动端。这次发布再次表明，手机正在从查看最终结果的终端，转变为软件 agent 的控制界面。

Aditya Agarwal says the quality bar for software has risen sharply: a vertical SaaS or agent-for-X company now needs either a truly differentiated vision or extraordinary growth. At the same time, he reports strong investor interest in frontier technologies such as robotics, drones, and semiconductors, suggesting that capital is looking beyond increasingly crowded software categories.

创业者 Aditya Agarwal 表示，软件产品的优秀门槛已经显著提高：一家 vertical SaaS 或 agent-for-X 公司，现在必须展现真正差异化的愿景，或者极其出色的增长速度。与此同时，他观察到投资者对机器人、无人机和半导体等 frontier technology 兴趣强烈，说明资本正在把视线投向日益拥挤的软件品类之外。

Anthropic introduced a Swift package that lets Apple developers call Claude through Apple's Foundation Models framework. Apps can keep fast, local work such as summarization and extraction on-device, then hand complex requests to Claude for multi-step reasoning, code generation, web search, and code execution. Typed Swift values produced through `@Generable` become clean inputs to Claude, while the package returns streaming output, tool calls, and structured responses to the same SwiftUI view. Anthropic says support will be available the next day across iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27, using an Anthropic API key.

Anthropic 推出一个 Swift package，让 Apple 开发者可以通过 Apple 的 Foundation Models framework 调用 Claude。应用可以把 summarization、extraction 等快速本地任务留在设备端，再把 multi-step reasoning、code generation、web search 和代码执行等复杂请求交给 Claude。通过 `@Generable` 生成的 typed Swift values 会成为 Claude 的干净输入，而 package 则把 streaming output、tool calls 和 structured responses 返回到同一个 SwiftUI view。Anthropic 表示，这项支持将在次日面向 iOS 27、iPadOS 27、macOS 27、visionOS 27 和 watchOS 27 提供，并使用 Anthropic API key。

The Takeaway: Wired's Kevin Kelly argues that AI is a "fifty year overnight success," and that today's systems are useful even though we still understand intelligence as poorly as early scientists understood electricity. He suspects intelligence is not a single element but a compound assembled from cognitive components we have not yet identified; AI may therefore fill a broad space with many engineered kinds of thinking rather than converge on one universal intelligence. Kelly's broader method is to visit the moving technological frontier, then return to history and hands-on work for perspective. His errors on VR and blockchain made him skeptical of confident timelines: physical robotics, in particular, may advance much more slowly than software AI because the human body combines extraordinary computation and mechanical efficiency. The durable lesson for builders is to watch for enabling breakthroughs from unexpected directions while staying humble about adoption speed.

核心结论： Wired 的 Kevin Kelly 认为，AI 是一场“历时五十年的一夜成功”，而且即使我们对 intelligence 的理解仍像早期科学家理解电一样有限，今天的系统也已经足够有用。他猜测 intelligence 不是单一元素，而是由尚未识别的认知组件组成的 compound；因此，AI 可能会在广阔的可能性空间中形成许多经过工程设计的思维类型，而不是收敛为一种 universal intelligence。Kelly 的方法是短暂进入不断移动的技术前沿，再回到历史研究和动手实践中校准视角。他在 VR 和 blockchain 上的误判让他不再轻信明确的时间表：尤其是 physical robotics，由于人体同时具备惊人的计算能力和机械效率，其进展可能远慢于软件 AI。对 builder 而言，持久的启示是关注来自意外方向的 enabling breakthrough，同时对采用速度保持谦逊。