Podcast Atlas

一个中国 FDE 的光环、落差与「救火」日常｜S10E27

Source note Episode guide Original audio Topics: Technology

Summary

This What’s Next｜科技早知道 episode uses 申越’s field experience to strip the glamour from FDE work in Chinese enterprise AI projects. The source argues that the hardest work is often not model or platform engineering but customer-side authority, business-team access, workflow discovery, document cleanup, acceptance criteria, and the vendor-side task of keeping a troubled project alive.

The episode sharpens Chinese-Style FDE / 中国式 FDE by contrasting a state-owned enterprise rescue case with a private-enterprise ROI case. In the state-owned project, the work could not move until a responsible customer-side owner could mobilize business departments; in the private company, the pressure shifted toward measurable usage, token consumption, labor-saving assumptions, and whether AI could fix a deeper lack of customers.

Key Claims

  • FDE is described as customer-site work that understands business, maps workflows, and brings AI into production, but 申越 says the domestic reality often looks more like vendor-side communication, coordination, and firefighting than a high-status technical-expert narrative.
  • In the state-owned enterprise case, the AI platform project nearly failed because the temporary IT owner could not mobilize business teams, the demos were based on imagined scenarios, and a leadership change removed earlier acceptance.
  • The turning point came when a new customer-side owner could contact eight or nine business departments and let the project team collect real materials, discuss workflows, and build an MVP around a department with heavy research and writing needs.
  • 申越 ranks FDE capabilities as communication first, industry experience second, AI cognition third, and technical implementation method fourth, while many job descriptions invert that order by foregrounding model, frontend, backend, fine-tuning, and parameter terms.
  • Private-enterprise projects apply more direct ROI audit: daily active usage, token spend, business relevance, theoretical headcount savings, revenue growth, and cost reduction all become acceptance pressure.
  • The source warns that AI cannot solve every business problem. If a store has no customers, adding AI agents may not fix demand, traffic, or business-model weakness.
  • Production AI depends on unglamorous preparation: data governance, internal-system integration, knowledge-base cleanup, machine-readable documents, clear contract-review rules, customer confirmation, and realistic accuracy expectations.
  • The legal contract-review case shows the expectation gap: business teams may expect AI to infer internal rules from law or competitor products while refusing to provide rules, yet still demand near-perfect accuracy.
  • The source treats current model and agent technology as early and probabilistic; large-scale enterprise production requires expectation management, human cooperation, and accuracy controls before deployment can be trusted.

Key Quotes

“第一是沟通，第二是行业经验” - 申越’s capability ordering for practical FDE work.

“有了 AI 你还是没客户” - the source’s private-enterprise demand-side caution.

“沈老师来了之后，这个项目终于有了转机” - the state-owned enterprise project leader’s later appraisal of the rescue work.

Connections

Contradictions

  • No direct contradiction found with existing wiki content.
  • The source reinforces Chinese-Style FDE / 中国式 FDE while qualifying more structured or glamorous FDE narratives: in practice, an FDE can be a low-power vendor-side actor unless the customer appoints someone who can mobilize business teams and define acceptance.
  • The source also qualifies AI automation optimism by showing that data cleanup, internal rules, human cooperation, and business demand may be the binding constraints rather than model access.