Unique Research · Unique Viewpoint · March 21, 2024
Background
On January 24, 2024, Unique Capital and partners held the 2024 AIGC Application Development Summit in Shanghai (~1,000 attendees). Edited from the guest dialogue.
Guests
- Chen Ran — Founder & CEO, OpenCSG (开放传神)
- Duan Xiaolei — GM, Global Startup Ecosystem Development, Huawei Cloud
Duan frames Huawei Cloud startup programs (vouchers up to ¥1M/startup, accelerator, go-global + AI focus) and introduces OpenCSG as an ecosystem member.
What does “OpenCSG” mean?
Chen: Company founded 2023; he has ~20 years in open source. Ex-Mesosphere colleagues went to OpenAI / Hugging Face — crystalizing certainty vs uncertainty. Certainty: AIGC/LLMs replace knowledge labor. Pain: models matter but don’t land. Mission: cost-down / efficiency-up on compute, data, human productivity.
- Open — open-source direction
- C = Converge — compute as priority; China still catching up vs U.S. lead → mix high/mid/low domestic + foreign GPUs on demand. Bilateral imbalance: sellers with idle capacity vs buyers who can’t find it.
- S — human efficiency: BAT wage inflation hurt digital transformation; use LLMs for code/dev/test/delivery.
Provocative claim: foundation models themselves aren’t valuable to commercialize — like Linux kernel; distributions + ecosystems monetize. Models should be open-sourced to customers; customer data IP means trained models belong to customers. OpenCSG: open, cost-down LLMs adapted for everyone.
Who are the most active users?
Chen: Five pillars — model, datasets, code sets, ecosystem, compute — ecosystem is most critical. Advise big tech: pick a lane, don’t do everything. OpenCSG ships a Hugging Face–like hub that is online + offline: ToB needs offline HF so data stays private. Future enterprise assets = datasets + code sets + model sets. Target users start as DevOps/cloud-native engineers (~30–40M open-source-literate pool as transition) evolving into AI engineers = prompt + cloud-native. Without toolchains/engines, new AI apps won’t proliferate. Mid-state users: technical builders; future-state: non-coders via natural language creating agents/native apps — model-to-model communication era.
What must a person/company “pay” for their own model?
Chen: Human socialization ≈ pretrain/prompt/fine-tune (parents → school → college). Differentiation is data. People are information carriers; DBs exist to store valuable corp data; valuable models are data-trained — orgs will have many models. Moment arrives when compute/data/code costs fall enough. Scenarios: personal assistants (health, education, travel) as agents talking to other models. Who replaces you isn’t the model — it’s people who use models well. U.S. ecosystem more complete; China’s still forming.
Personal data scattered across apps — how become my asset?
Chen: Huge commercial opportunity in helping individuals collect/produce/trade lawful datasets. National Data Administration is step one. Harder: norms defining ownership (individual / group / public / government). Large firms won’t share proprietary data (no moat without it). Eco partners must resonate → co-create → business models. After factor rules, private vs public boundaries clarify — private data won’t train public general models.
Hundreds of models launched since 2023 — their future?
Chen: Uncertainty = business model. “Any startup not aiming at a business model is playing rogue.” Models themselves may not commercialize, but around them: software subscription, dataset delivery, services, commercial distribution. Model = brain; agents = hands/feet. Knowledge monetizes via data; agents monetize by running.
OpenCSG’s own business model
Online: lowest cost/barrier for developers; compute revenue-share with partners (incl. Huawei). Offline: model hub like GitLab + commercial software subscription — sell software, not the model. Goal: lowest cost/barrier for offline individuals/freelancers to use LLMs.
3–5 year ambition
Pre-2023: China apps / U.S. infra; now China must build its own infra. 14th Five-Year Plan elevates open source. OpenCSG aims to be China’s first IPO in open LLM tooling — China’s Hugging Face for developers.
4 CN images; attach on ingest.
— END —