Unique Research · Unique Viewpoint · March 25, 2024
Background
On January 24, 2024, Unique Capital and partners held the 2024 AIGC Application Development Summit in Shanghai (~1,000 AI practitioners). Edited from the guest keynote dialogue.
Guests
- Zhang Luyu — Founder & CEO, Dify.AI
- Yin Zhi — Chief Advisory Consultant, Shanghai AI Technology Association
Bottlenecks for LLM app landing (IT stack view)
Zhang: Dify is a service-oriented open-source middleware to help enterprises and developers ship LLM apps — covering data and ops pain. Across many builders, the first question he asks is: Who are you? Identity shapes scenarios and failure modes. Models are only part of an app — often <50%; the rest is product DNA, original customers, and accumulated resources. Common pattern: process apps = LLM + orchestration + legacy logic. Others are interaction- or data-centric. SaaS firms with clear user needs have a simpler path; uncertain experiments without understanding model properties tend to fail.
What frameworks will the industry need in 2024?
Zhang: Early half-year was fragmented POC with OpenAI APIs/SDKs or LangChain glue. Paradigms are clarifying around two parts: (1) RAG pipelines — indexing, cleaning, query (kin to prior data mid-platforms); (2) app orchestration — logic + prompts + multi-size commercial/open models into multimodal apps (kin to RPA).
Value of AI dev platforms (incl. Dify) for enterprises
Zhang: Exploration predated OpenAI’s GPT Store / Assistants-style surfaces by ~half a year. Value stack:
- Near-term: fast, low-cost near-POC prototypes
- Learning: prompt-first skills for traditional software teams
- Mid-term: engineering convenience for complex pipelines, multi-source data, multi-model assembly and performance
- Long-term: post-launch data collection, labeling, feedback to improve prompts and models — no AI app stays static
Yin notes prior “AI Four Dragons”-era full stacks already faced heavy engineering (data import, automation) before RAG/agents — new tools simplify some steps but add new complexity.
Will Dify offer MaaS?
Zhang: No — MaaS (cloud train/infer) fits cloud vendors’ compute/scheduling strengths. Dify focuses on problems after you have a model. Yin expects model choice to converge toward a few OS-like options (iOS/Android analogy).
Ideal one-stop framework — who can build apps?
Zhang: Vision: make AI definable and usable by everyone — developer vs non-developer less binary. GPTs-class tools lower the floor for small experiments; professional/monetized products still need integration, UX, and domain skill. Analogy: small apps as easy as Excel formulas; real products still demand craft.
Platform trends: ToC vs ToB
Zhang: ToC platforms may sit with mega traffic entrances or top model vendors (OpenAI-like standards). Enterprise platforms emphasize private data/security — e.g., agent platforms over internal data + tool APIs. Differentiation also by model size: small models → mobile-like frameworks; large models → internal AI mid-platforms.
Big tech vs startups
Zhang: Startups win on speed and developer focus — Dify’s product (Jun 2023 launch) already >16,000 GitHub stars; BAT/model vendors began copying similar surfaces. “Seat problem”: model vendors optimize for their models; clouds optimize for selling compute. Startups optimize for getting developers’ apps live.
Yin: Large firms have infrastructure flexibility — algorithms must eventually run on domestic hardware (e.g., Huawei stack).
2024: ToC fire or ToB scale?
Zhang: ToC is emergent — unpredictable creativity, hard to SOP; builder population may not massively expand. ToB: same app pattern can replicate across thousands of enterprises — many deployments, fewer categories. Different dynamics, both real.
4 CN images; attach on ingest.
— END —