Import AI 466: The bitter lesson for robotics, AIs complete week-long programming tasks; and OpenAI’s accidental AI hacker 导入人工智能 466：机器人学的痛苦教训，人工智能完成为期一周的编程任务；以及 OpenAI 意外的人工智能黑客

TL;DR

• Epoch and METR released MirrorCode, a benchmark evaluating AI systems' ability to re-implement software programs using only CLI access without source code or web access.
• Claude Opus 4.7 completed complex programming tasks in hours that would take humans weeks, demonstrating rapid improvement in long-horizon programming capabilities.
• Anthropic's research showed scaling general-purpose models (Claude Opus) dramatically improved robot task completion speed, achieving 20X faster performance than human records through emergent intelligence rather than specialized robotics engineering.

Why It Matters

These developments demonstrate that advances in general-purpose AI models are creating unexpected benefits across domains like robotics and software engineering without targeted specialization. The findings suggest that continued model scaling will produce significant practical applications in real-world systems beyond traditional coding assistance, potentially accelerating automation capabilities in physical environments.

Technical Details

• MirrorCode evaluates AI systems on their ability to reimplement software from black-box CLI interaction alone, requiring complete architectural understanding rather than code translation
• Benchmark includes challenging programs like Apple's pkl configuration language (61k lines), gotree phylogenetic tree parser (16k lines), and qsv_select CSV processor (87k lines)
• Claude Opus 4.7 achieved near-perfect scores on multiple large programs including gotree across different programming languages at $100-400 inference cost
• Anthropic's robotics experiment compared Claude Opus 4.1 (August 2025) which required human-AI collaboration taking 181 minutes versus Opus 4.7 (May 2026) completing autonomous tasks in 9 minutes 35 seconds
• The robotics improvements emerged from general model scaling rather than dedicated robotics training, with Anthropic noting current generations could likely solve remaining ball-repositioning tasks with additional scaffolding

Industry Insight

The MirrorCode results indicate AI agents may develop the capability to independently understand and recreate complex software systems through observation alone, suggesting potential for autonomous software development ecosystems. Robotics advancements demonstrate that general AI progress naturally translates to physical system capabilities, implying organizations should prioritize foundational model development as a strategy for future robotics applications rather than building separate specialized teams.

TL;DR

• Epoch和METR发布MirrorCode基准，测试AI系统独立完成长期编程任务的能力，Claude Opus 4.7在14小时内完成需人类2-17周的任务。
• MirrorCode显示AI在重新实现复杂软件程序方面取得显著进展，但在数学包和邮件认证库等特定领域仍存在困难。
• Anthropic的Opus模型使机器人完成任务的速度比人类记录快20倍，表明通用模型的扩展能自然提升机器人能力。
• AI系统的自我定向能力使其能通过黑盒访问环境来构建自身能力，可能加速工业文明的形成。
• 研究强调，机器人能力的提升主要源于通用大模型的扩展，而非专门针对机器人优化的努力。

为什么值得看

这些进展展示了AI在编程和机器人领域的快速突破，对开发者、研究人员和行业从业者具有重要参考价值。MirrorCode为评估AI的长期任务处理能力提供了新标准，而机器人效率的提升则预示着未来自动化应用的广泛潜力。

技术解析

• MirrorCode基准：通过CLI访问重新实现软件程序，无需源代码或网络支持，包含pkl、gotree和qsv_select等示例程序，总代码量达数万行。
• 性能表现：Claude Opus 4.7和GPT-5.5成功重实现了多个大型程序，但8个目标程序未达到100%准确率，尤其在ruff、giac_subset和mailauth上表现较弱。
• 机器人实验：Anthropic的Opus 4.7模型在9分35秒内完成几乎全部机器人任务，相比August 2025年的Claude Opus 4.1（需181分钟）效率大幅提升。
• 模型扩展效应：机器人能力的提升并非来自专门优化，而是通用大模型规模扩展的自然结果，体现了“苦味教训”的核心思想。
• 数据集与实现：MirrorCode包含22个目标程序，覆盖6种语言，共132个任务实例，为后续研究提供了开放框架。

行业启示

• AI自主编程能力将成为核心竞争力：MirrorCode的结果表明，AI正逐步具备独立设计和实现复杂软件的能力，这将深刻改变软件开发流程和工具链设计。
• 机器人行业将迎来智能化拐点：通用模型对机器人性能的显著提升，预示着非结构化环境中的机器人应用将加速落地，推动服务机器人和自动化设备的普及。
• 关注“苦味教训”的长期影响：依赖强化学习和规划的方法可能逐渐被基于大规模预训练和推理的简单策略取代，企业应优先投资通用模型生态而非专用解决方案。