IndustryMulti-source

John Schulman 与 Mark Chen 对谈:模型的"研究品味"能否通过 RLHF 学会

Why we picked it

OpenAI 和 Thinking Machines Lab 的两位核心人物同台聊 AI 自动做研究这件事。Chen 直说模型还只会调参数不会提想法,Schulman 给出了用 RLHF 培养研究品味的具体方案。做 AI 研究或想了解前沿实验室怎么思考 RSI 的人都该看看。

a16z 播客 Working Theory 第一期上线,Sarah Wang 与 OpenAI 首席研究官 Mark Chen、Thinking Machines Lab 联合创始人 John Schulman 讨论递归自我改进(RSI)。Schulman 认为"研究品味"指判断哪些想法长期重要、能超越原始论文并与其它想法组合的能力。Chen 指出当前模型在品味上表现不佳,只是擅长超参数优化等可调参数的深度调优。

Image · a16z
Body

Schulman 提出两条培养路径:一是从领域历史文献中吸收前人经验,二是用 RLHF 从研究者那里收集"哪些想法好、哪些不好"的数据,把人类直觉烘焙进模型。节目还讨论了 RSI 的定义、距离完全实现还有多远,以及为什么 AI 可能是解决对齐问题的工具。

Original · a16z

John Schulman and Mark Chen on what research taste is, and how models might learn it: John: "Research taste is about knowing which ideas are important in the long run, which ideas really catch on and end up going beyond whatever paper they first appear in, but being general and being composable with other ideas." Mark: "There's taste, which is how you come up with the idea in the first place. The models are still not phenomenally good at taste. They go very deep into things like hyperparameter optimization, essentially tuning all the little knobs that you can tune." John: "A lot of intuition just comes from personal experience doing research, but to get good at a certain research field, you also have to read a lot of the past work in that field, and you try to absorb the lessons from all this historical work." "Companies will probably also just do RL from human feedback to collect data from researchers on which proposed ideas are good and which aren't. And that's kind of like absorbing whatever human intuitions the human researchers have developed and trying to bake that into the models." @johnschulman2 @markchen90 @sarahdingwang

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Industry" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include John Schulman, Mark Chen, OpenAI, Thinking Machines Lab.

Source attribution: curated from a16z (@a16z). Cross-source corroboration: 11 independent sources have covered this development (including 向阳乔木, IT之家, OpenAI Blog), confirming industry-wide signal strength.