skyfallai 团队称 LLM 不适合商业决策，改做预测企业变化的世界模型
Maluuba 那帮人出来创业了，说 LLM 规划能力不行，改做能推演公司决策后果的世界模型，思路挺不一样。
由 Maluuba 原班人马创立的 @skyfallai 正在构建预测公司决策后果的模型。其创始人认为 LLM 过去 5 年的进步主要靠数据、算力和模型规模堆叠，但这类模型在长时程规划上表现差、依赖海量数据、且在现实条件变化时容易失效。他们选择用世界模型学习因果关系，让智能体能在仿真中先推演一个动作的后果再执行，对应的问题是'这个决策六个月后业务会变成什么样'。
A language model is trained to produce the next good response.
Running a company asks a different question: what does the business look like six months after this decision?
@skyfallai, founded by the team behind Maluuba, is building models that predict how a company changes after each decision.
Its founders argue that LLMs, which have advanced for 5 years mostly by adding data, compute and model size, are the wrong foundation for that job.
Their complaint is specific: current models plan poorly over long horizons, need huge amounts of data and break down when real-world conditions shift.
A world model attacks those weaknesses by learning cause and effect, so an agent can rehearse an action in simulation before taking it.
Below is a great read to learn more about the society post-automation.
