Daily Digest · 2026-07-30

今日精读

今日焦点：OpenAI 揭示 GPT-5.6 Sol 在 ARC-AGI-3 低分的根本原因——评估 harness 限制，并给出 API 最佳实践；同时宣布向全球研究人员免费开放前沿模型。Anthropic 则补全了密码学攻击的技术细节，并发布可复现基准。

2026-07-29 至 2026-07-30（约24小时）8 条40 推文信号强

01 关注动态

OpenAI 揭秘 GPT-5.6 Sol 在 ARC-AGI-3 低分原因：评估 harness 限制，并推荐 retain reasoning 与 compaction

“GPT-5.6 Sol has been used to solve open problems in mathematics. So why was it struggling with ARC-AGI-3, a benchmark of 2D puzzle games? We investigated. The harness was not letting it remember what it had learned. We found that enabling two API settings tripled our scores with 6x fewer output tokens.”

实用点

API 开发者应使用 Responses API（而非遗留 Chat Completions）、开启 retain reasoning 与 compaction 来最大化 agent 持续推理能力。这对部署 GPT-5.6 系列 agent 的生产级系统有直接指导意义。

@OpenAI 原文1

02 关注动态

OpenAI 公布 GPT-5.6 Sol 自我优化结果：服务成本降低 20%，推测解码效率提升 15%+

“After deployment, we applied GPT-5.6 Sol to advance the frontier of efficiency by making itself more efficient to run. The results: 20% lower serving costs from production GPU kernel improvements. 15%+ better token-generation efficiency from improved speculative decoding.”

实用点

展示模型自我优化潜力，考虑部署 GPT-5.6 的团队可预期成本下降与 token 生成速度提升，尤其在大量 API 调用的场景下收益明显。

@OpenAI 原文1

03 行业趋势

OpenAI 启动 ChatGPT for Academic Researchers：免费向科研人员提供 GPT-5.6 系列，目标十万研究员

“We’re giving scientists, mathematicians, and engineers free access to our frontier models—starting with 10,000 researchers and expanding to 100,000 through 2027. ChatGPT for Academic Researchers is built to accelerate discovery across disciplines.”

实用点

研究人员、学术机构可申请免费访问 GPT-5.6 系列模型，包含业务级隐私保护（数据不用于训练，最多邀请 4 位协作者）。对于从事计算科学、数学、工程等研究的开发者，这是获取前沿模型成本极低的入口。

@OpenAI 原文1

04 大佬官方

Anthropic 发布 HAWK 与 AES 密码学攻击完整技术论文及 CryptanalysisBench 基准（新增细节）

“Full technical details of both attacks are provided in our new papers: On HAWK... On AES... And the associated model chain-of-thought for AES.” “We also worked with academics at ETH Zurich, Tel Aviv University, and the University of Haifa to build CryptanalysisBench, a benchmark for studying LLMs’ cryptanalysis abilities.”

实用点

安全研究者可查阅详细论文了解 Claude 发现密码学弱点的完整方法；CryptanalysisBench 可用于评估 LLM 的密码分析能力，填补了此前仅有概述的空白。

@AnthropicAI 原文1

05 行业趋势

Anthropic 签署支持 AI 发展节奏控制的请愿，CEO 及多位高层共同署名

“We support this petition, signed by our CEO, several co-founders, and senior staff. Our own research on recursive self-improvement...points to the need for tools to deliberately pace the frontier of AI development so society can prepare.”

实用点

反映主要前沿实验室对 AI 安全治理的共识姿态。开发者可关注未来可能出现的监管工具（如 pacing tools），以及企业内部治理要求的演变。

@AnthropicAI 原文1

06 工具更新

企业开发者演示用 Copilot + Microsoft Fabric 构建 ROIC 智能应用，展示完整治理流程

“I took the PDF that Brian Nowak at Morgan Stanley put together for Hyperscale ROIC this week and used Copilot code (coming in our new superapp) with a single prompt + skill (/drill-me) to create the plan, then used autopilot in auto to create the full app... And the best part is that all the artifacts are in my enterprise environment... everything is under Agent 365 IT/Sec/FinOps control!”

实用点

展示了在企业治理框架（IT、安全、财务运营）下，用自然语言构建完整数据应用的实际流程，适合企业开发者参考将 AI 代码生成纳入标准开发管道。

@unknown 原文1

07 大佬官方

Pieter Levels 宣布 "SaaS is dead" 并推出 "Can I Vibecode It?" 网站，提供即用 Prompt

“SaaS is dead. Most subscriptions are just a prompt. Can I Vibecode It? gives you the prompts. Copy. Paste. Cancel.”

实用点

独立开发者可将复杂 SaaS 服务转化为可复用的 AI prompt 模板，从而快速替换现有订阅。该网站提供了具体 prompt 集合，可直接试用获取灵感。

@levelsio 原文1

08 其他

GitHub 启动残障开发者可访问性调查，截止 8 月 31 日

“GitHub wants to learn more about how developers with disabilities contribute to software and how accessibility technology helps them do their best work. We welcome you to fill out our survey, open until August --- _Source: X via free cookie CLI + DeepSeek. Hosted on GitHub Pages._