THE BLOG / RESEARCH & PRACTICE
Research, experiments and lessons from building AI systems. How models meet people, products and the real world.
AI SYSTEMS · HUMAN JUDGMENT · ENGINEERINGThe 80,000-game model is live. Its bounded near-best policy reached 190/200, 462/500, and 915/1000 across fixed, validation, and a fresh final set. This report records training, failed experiments, feature expansion, and the policy update.
Read the noteThis design note examines an unimplemented RPA idea: using AI to interpret text and a checkerboard aligned with pointer-motion units to express click or drag coordinates. It then explains why third-party production CAPTCHAs should not be treated as ordinary UI automation, and reframes the idea as a safe workflow built around official APIs, test credentials, human review, coordinate calibration, confidence gates, and audit logs.
This essay argues that domain-specific world models will probably mature before AGI that can act safely in open physical environments, while a complete universal world model may not arrive as a separate system before AGI. The practical safety unit is an assurance stack combining prediction, scenario generation, constraints, monitoring, and staged deployment.
We compare One-shot, pre-task clarification (PPN), and iterative disclosure and repair (IDR): PPN sits between the other two in cost, but its quality advantage did not replicate across models.
Instead of feeding AI more examples of innovation, we could place it in a controlled scientific environment with no given answer and test whether it can move from observations to laws.
THE BLOG / RESEARCH & PRACTICE
记录构建 AI 系统过程中的研究、实验与经验，以及模型与人、产品和现实世界相遇时的问题。
AI SYSTEMS · HUMAN JUDGMENT · ENGINEERING80,000 局模型已上线；近优随机策略在固定、验证和新最终评估中分别达到 190/200、462/500、915/1000。本文记录训练流程、失败实验、特征扩展及策略更新。
阅读全文本文记录一个尚未实施的 RPA 设计设想：文字界面可由 AI 辅助理解，点击或拖动界面可用与鼠标步长对齐的棋盘网格表达坐标。文章进一步解释为什么生产环境中的第三方 CAPTCHA 不应被当作普通 UI 自动化，并将该设想收敛为官方 API、测试密钥、人工介入、坐标校准、置信度门控和审计日志组成的安全工作流。
本文提出一个经过收敛的判断：领域级 world model 大概率会先于面向开放现实世界的 AGI 成熟，但完整通用 world model 不一定独立早于 AGI。真正的安全关键不是单个模型，而是 world model、场景生成、约束验证和分阶段部署组成的 assurance stack。
我们比较 One-shot、执行前澄清（PPN）和迭代式对话修复（IDR）：PPN 的成本位于两者之间，但质量优势没有跨模型复现。
与其让 AI 背更多“创新案例”，不如构造一个没有现成答案、但允许观察和实验的科学环境，测试它能否从现象走到规律。