Gary Marcus 与 Nate Silver 隔空争论 LLM 是否只是"随机鹦鹉"
Gary Marcus 和 Nate Silver 打起嘴仗了，吵的是现在的大模型到底还算不算"随机鹦鹉"，两边观点都挺典型，看看你站谁。
Gary Marcus 在 X 上批评 Nate Silver 等人混淆了纯 LLM 与带工具、代码解释器的混合系统的区别。Marcus 认为早期纯 LLM 大致符合"随机鹦鹉"的描述，如今神经符号化的混合系统工作方式不同，但这不能证明纯扩展路线是对的。Nate Silver 回应称，让 Astra 或 Fable 调试一个 1 万行程序的表现远超"随机鹦鹉"的说法。Marcus 补充称目前进展多限于可验证领域，可靠的通用智能尚未到来。
it’s a shame to see someone like @NateSilver538 does not understand a basic, fundamental technical distinction. ( @tristanharris and Kevin Roose missed it the other day, too.) pure LLMs were (very roughly and certainly not exactly) like stochastic parrots. harnesses, tools, code interpreters heralded the spread of neurosymbolic AI. the new systems are hybrids; they work differently. the somewhat weak metaphor is much weaker with them. but the fact that they are less like stochastic parrots doesn’t mean the original systems weren’t, nor that pure scaling worked. it means that neurosymbolic AI was an important next step. this still doesn’t mean we have AGI; most of the progress works best in verifiable domains; truly reliable general intelligence has certainly not yet arrived. (my 2020 article Next Decade in AI foresaw all of this, and what’s still missing, including world models, which i called cognitive models there.) Nate Silver @NateSilver538 OK then give Astra or Fable a 10,000-line program, ask it to debug it, ask it to understand what's going on, ask it to make revisions. It's certainly not perfect. I'm bullish on human expertise fwiw. But it sure as hell is way more than a "stochastic parrot." 🔗 View Quoted Tweet 💬 10 🔄 5 ❤️ 26 👀 3112 📊 10 ⚡