Gary Marcus又开炮了，这次用国际象棋例子说明LLM缺乏常识，连简单规则都学不会，你还敢让它管战争？
Gary Marcus引用卡斯帕罗夫的推文，指出现成的LLM在国际象棋中经常做出非法走法，说明它们缺乏真正的世界模型。例如，2023年多次观察到LLM无法准确理解简单的规则集。Marcus认为，如果一个系统连国际象棋规则都无法可靠遵循，就难以相信它能处理需要上下文和常识的通用任务。他批评那些希望用这类系统指挥战争的想法是疯狂的。
Exactly, @Kasparov63 One of my favorite examples of LLMs lacking world models is the failure of off...
Exactly, @Kasparov63 One of my favorite examples of LLMs lacking world models is the failure of off the shelf LLMs to follow the rules of chess. Insane to think people want systems that can’t learn chess without massive data (except by relying on purpose-built external tools) to run wars. Garry Kasparov @Kasparov63 Perhaps! But do LLMs still 'cheat' by making illegal moves? Saw this many times just last year. If they can't accurately interpret a ruleset as simple as chess, how can they accurately generalize, which takes context and what we call common sense? 🔗 View Quoted Tweet 💬 3 🔄 2 ❤️ 14 👀 1907 📊 4 ⚡