精选理由
Gary Marcus直言LLM的五大局限，解释AGI需要什么，和当前技术有何不同。
Gary Marcus指出，LLM在搜索和多种人类任务上表现出色，但缺乏心智理论和世界模型。LLM无法持续计算，容易偏离轨道，训练和运营成本极高。AGI需要更多技术和模型，特别是神经认知方面的，AI工作流和芯片将向异构化发展。
原文 · Gary Marcus
💯
💯 Jason Pontin @jason_pontin Well, to express myself more soberly… it turns out that LLMs, trained on all the text in the world are phenomenally useful—better at search, and generative for a surprising variety of human tasks—but LLMs lack a theory of mind and world, cannot continuously compute, go off the rails, and are ruinously expensive to train and operate. AGI will require further techniques and models, many of them neuro-cognitive, and AI workflows and chips will evolve to be heterogeneous. 🔗 View Quoted Tweet 💬 1 🔄 0 ❤️ 4 👀 1195 📊 1 ⚡