Gary Marcus 批评将 2022 年 LLM 与 2026 年神经符号系统混为一谈的论调
Gary Marcus 这次难得没全盘否定 AI，他说 2026 年的系统加了符号验证早不是当年的“随机鹦鹉”，还提到万级智能体 88 小时攻 Navier-Stokes，观点有点意外。
Gary Marcus 发文批评一些人把 2022 年的纯 LLM 与 2026 年结合符号框架、代码解释器与符号验证的系统混为一谈，以此做修辞性论点。他引用 Emad Mostaque 关于 AI 用 10000 个智能体在 88 小时内求解 Navier-Stokes 证明的说法，指出“随机鹦鹉”这一比喻只适用于当时的旧技术。Marcus 认为这种说法低估了后续创新和神经符号 AI 的贡献，并批评多数谈论 AI 的人并不了解底层技术原理。
We need a name for the kind of people that conflate 2022 pure LLMs that were in fact a tiny bit like stochastic parrots (not my phrase, and not one I ever endorsed) with 2026 systems that couple those LLMs together with symbolic harnesses, code interpreters, and symbolic verification, typically to make a slippery rhetorical point. Aside from being unfair to people who made a metaphor that applied (to some degree) to the technology of that time, it disrespects the innovations that were subsequently made and disrespects the critical contributions that neurosymbolic AI made to progress. What I hate most about this site is that so many people talk about AI yet few take any time to understand how the different underlying technologies work. Most prefer to make cheap rhetorical points instead. Which is exactly why we are in the mess we are in. Rohan Paul @rohanpaul_ai Emad Mostaque (founder of Stability AI) on Navier-Stokes proof by AI. "Do you really think there's anything you can do cognitively that this AI can't do? 88 hours with 10,000 AI agents solved it, thats 100 years of top level mathematician work. Think about that. Before this week, you could say, "Ah, it's a stochastic parrot. It can only learn from its distribution." And then you have literally a millennium prize problem. There's only 7 of them. Humans have figured out one. and a swarm of 10,000 AIS over 88 hours solves it. You can't say that's in the training data" --- (Full video on 'Risk Takers' Youtube channel, linke in comment) Your browser does not support the video tag. 🔗 View on Twitter 🔗 View Quoted Tweet 💬 10 🔄 0 ❤️ 34 👀 4739 📊 11 ⚡