- Xu attributed continued AI progress to pretraining compute, post-training and inference, and agent design.
- She said higher verification costs and lower error tolerance complicate adoption beyond coding.
A shifting implementation bottleneck
Moonshot AI's head of financial industry and ecosystem, Xu Tiantian, said businesses' AI adoption challenges are moving from asking questions and organizing information toward building systems. She spoke at the seventeenth Caixin Summit on November 12, according to Sina Finance's same-day report.
Xu argued that corporate institutional design is advancing more slowly than models. She also said agents can work independently for longer periods and are moving into deeper workflows. Those statements describe her assessment of the industry, rather than a measured deployment result reported by the outlet.
Verification costs remain a constraint
Xu identified three continuing sources of AI progress: pretraining compute, post-training and inference, and increasingly complex agent design. Their combined growth, she said, supports faster iteration.
She contrasted coding, where checking outputs can be relatively quick and inexpensive, with fields such as finance, law and healthcare. Higher verification costs, low tolerance for errors and more complex systems create difficulties in those industries, she said. The report says Moonshot is exploring these questions with industry participants; it does not identify a new product, customer deployment or quantified improvement.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01


