When AI Can Do Everything, Do We Still Know What We Want?

Ryo Lu has released his full talk from AI Engineer Shanghai, and the title is a question: When we can do anything, do we still know what we want?

The talk video has Chinese and English subtitles, and all the slides were made by hand in Keynote. Put this detail together with the question, and it's a bit interesting. Tools lower the barrier to production, while doing things by hand becomes a choice.

Over the past year, the AI toolchain has lowered the barrier of 'can it be done.' Models can write code, agents can connect to tools, and the path from prototype to product has become shorter. Once implementation gets faster, the bottleneck does not disappear; it just moves. More teams can quickly build features, yet they cannot clearly say what to build, what level of completion counts as right, or how to judge whether it was worth it after it's done.

This is what makes the talk's question worth examining. Expanding capability does not automatically bring direction. The stronger the tools, the more expensive intent becomes. Being able to generate ten options does not mean knowing which one to choose. Being able to build a prototype in a day does not mean knowing whether it should be built.

For AI practitioners, this shift is very concrete: requirements documents, evaluation criteria, product boundaries, and stopping conditions—the parts that don't look sexy—determine what the tools ultimately produce. The same goes for ordinary users. When AI can write, draw, search, and make spreadsheets for you, the real question to answer is which task you want it to complete on your behalf, and what you are willing to bear for the result.

The video is below, with Chinese and English subtitles. Source: Ryo Lu's tweet.

Original tweet