I was making progress on that page but it didn’t feel like a Red Blob Games page. The page started out with tons of shell commands, and then showed lots of code. It felt like a page that only I would find useful. So I started over and designed a “concepts” page. In redesign 4 I focused on what effects I wanted, how SDF works, and how to use it to create those effects. I again reduced the scope by removing the implementation details. What I had already written, I moved to a separate (unpolished) page. And I never wrote a standalone downloadable project like I originally wanted.
在 AI 场景中，Apache Spark 凭借其强大的批处理能力与 Python 生态兼容性，广泛用于大模型训练前的数据清洗、特征工程与推理任务。而 Ray 因其低延迟、高并发特性，被 OpenAI 等头部机构用于分布式训练与强化学习。两者共同构成 Data + AI 的核心计算底座，支持从数据准备到模型推理的全流程高效执行。
。夫子对此有专业解读
用产品经理的心态对待咖啡，不断迭代好喝的咖啡。公众号：咖啡平方
其次，大模型的记忆能力有缺陷：大模型在训练时“记住”了大量知识，但训练完成后并不会在使用中持续学习、“记住“新知识；每次推理时，它只能依赖有限长度的上下文窗口来“记住”当前任务的信息（不同模型有不同上限，超过窗口的内容就会被遗忘），而无法像人一样自然地维持稳定、长期的个体记忆。但在真实业务中，我们需要机器智能有强大的记忆能力，比如一个AI老师，需要持续记住学生的学习历史、薄弱环节和偏好，才能在后续的讲解与练习中真正做到“因人施教”。