Zhao Chenyang: The Way We Evaluate Technology Is Turning Back the Clock on History

On a podcast, Zhao Chenyang discussed creative labor in the age of AGI and used a weighty phrase: "the collapse of rites and music" (li beng yue huai). Another phrase he used was "sorrow and joy intertwined" (bei xin jiao ji). Put together, the two point to the same thing: when technology changes fast enough, the yardstick formerly used to measure the value of labor can suddenly fail.

He is involved in developing the open-source inference system SGLang and has both research training and industry experience. The title of this conversation is "In Conversation with RadixArk's Zhao Chenyang: Open-Source Inference Engines, AI Evaluation and Trust—'The Collapse of Rites and Music' and 'Sorrow and Joy Intertwined.'" The image also includes the cover of a book titled "AI Writes the Draft, AI Reviews the Draft—But Who Takes Responsibility?" by Zhao Chenyang, with a blue box below reading "Bounded Rationality." Taken together, these details show that he cares not only about how fast models run, but also about how evaluation and trust are established.

One point Zhao Chenyang emphasizes is this: history textbooks say that the development from the nine-rank system to the imperial examination system was social progress. The nine-rank system selected people by family pedigree, while the imperial examination system used exams to broaden participation. Today, in his view, our evaluation system for science and technology is instead putting history in reverse. This judgment is jarring, but it is not an emotional outburst. It points to a concrete problem: if an evaluation system recognizes only metrics that can be scored and ranked, creative labor will be compressed into a single row in a spreadsheet. Work that can be entered into a form looks reliable; work that cannot looks suspect.

Creative labor is extremely fragile and extremely valuable. Zhao Chenyang says that reflection on the value of one's own labor and even the meaning of life is an ultimate question that cannot be avoided once you start thinking about it. This sounds unlike a technical discussion, but it and the evaluation system are two sides of the same coin. The evaluation system determines what is seen, what is rewarded, and what is eliminated. People doing creative work can hardly remain completely unaffected by this system.

The podcast was originally recorded over three hours, then re-recorded; five minutes were cut in between, along with a historical coincidence that Zhao Chenyang liked very much. He said that if anyone was interested, he could share it privately. In the replies, some people asked for the podcast link, while others urged him to talk about that historical coincidence. These details make the whole thing more than a piece of opinion output; it feels more like a scene that wasn't cleanly edited: some words were kept, and some were taken out.

Original tweet: https://x.com/GenAI_is_real/status/2115631869224198601