MIT's Liang Qiyao on the 'physics of AI' and her pessimism about research
In an interview with DeepTech published by MIT Technology Review China, MIT researcher Liang Qiyao says she is optimistic about AI and therefore pessimistic about scientific research, citing the concentration of money and talent in the field. She argues that large models show surprising capabilities that researchers still cannot explain, so current AI research retains an "alchemy"-like character in which people know which methods improve performance but not why. Trained in physics, she studies AI as a complex system using small "toy models", from compositional generalization in diffusion models to post-training of large language models with evolution strategies.
Liang argues that theorizing why training methods work is both a scientific question worth pursuing and a direction where academia can still compete with large companies.
“正因为对 AI 乐观,所以才对科研悲观。”采访中,梁琪瑶用这样一句看似矛盾的话,形容自己对 AI 与学术研究未来的判断。
几年前,她还在研究量子计算。作为一名接受物理学训练的研究者,她始终对世界的运行机制抱有好奇,希望通过理论与实验,解释那些尚未被理解的现象。2022 年底,ChatGPT 的出现,让她找到了一个新的研究对象——AI。
尽管大模型已经展现出令人惊讶的能力,但人们仍然很难解释,这些能力究竟是如何产生的。在她看来,如今的 AI 研究仍带有某种“炼金术”色彩。研究人员知道哪些方法能够提升模型性能,却未必清楚背后的原理。她希望改变这一点。