Key Takeaways from Asset Management Leaders on Large‑Model AI at the Bund Conference

Ant Group (VP & CTO of Wealth & Insurance Business Group)

Vertical (domain‑specific) models outperform generic large models; the main cost lies in training talent and building evaluation systems.

Integrating a suite of vertical models into scenarios yields better results than a single vertical model; the cost is tied to deep business expertise.

Business‑expert talent must be willing to contribute time and knowledge, which requires innovative operational incentives.

Guotai Haitong Securities (CIO)

Adopts a "1+N" model strategy: one foundational model plus N scenario‑specific models.

Comprehensive rollout includes model intent recognition for large models and output‑quality control for small models.

Innovation focuses on AI‑native applications (e.g., AI‑powered apps such as Junhong APP) and platform‑level capabilities.

Large‑model applications are shifting process digitisation toward full‑document digitisation.

Promotes a "slow‑thinking + fast‑response" paradigm for model usage.

Fortune Fund (CIO)

Advocates a C2C (code‑to‑code) development model that replaces the traditional D2C (document‑to‑code) workflow, moving toward semi‑structured markdown code.

Automation of CRUD operations under C2C can achieve near‑full development automation.

Automated testing efficiency improves by roughly 50%.

Enterprise‑level development efficiency is boosted by redefining standards, evaluation criteria, and reusable components within the C2C framework.

AI Playground (Self‑developed by a fund‑management tech team)

Provides a high‑performance, scalable platform for AI applications, supporting the four pillars of application engineering, model engineering, knowledge engineering, and compute engineering.

Investment‑Research Challenges (Shenwan Hongyuan Fund CIO)

Questioning the premise that "high‑quality corpora" directly improve investment decisions; the definition of high‑quality data is unclear.

Proposes a corpus‑selection flywheel for iterative self‑improvement independent of direct business contribution.

Mixture‑of‑Experts (MoE) models reduce cost but may compromise effectiveness.

Asset‑Management Model Path Exploration (Tianhong Fund AI Lead)

Emphasises multi‑step, human‑like, and divergent thinking to advance model capabilities.

Rebuilds search recall, process reasoning, and establishes evaluation systems to overcome key technical barriers.

Intelligent Investment‑Research Applications (Orient Securities Asset‑Management GM Assistant)

Builds two engines: data‑driven business decision support and intelligent data analysis/customer service.

Focuses on efficiency gains and enhanced user experience.

Notes that macro/industry analysis is feasible, but individual‑stock model analysis remains difficult.

Other Perspectives

Large models and autonomous agents are being explored for intelligent investment‑research scenarios (Industrial Securities).

Co‑existence of large and small models enables high‑impact use cases.

Model‑centric data collection, cleaning, and governance are essential for downstream applications.