Unique Research · Interview / Roundtable · May 9, 2024
From Unique Capital’s Jan 24, 2024 Shanghai AIGC summit (~1,000 attendees).
Panel
- Zhu Jiang — Founder & CEO, Jingying Tech (井英)
- Gao Yizhao — CEO, Zhizi Engine (智子引擎)
- Zhang Feibiao — Co-founder & CTO, Wujie AI (无界)
- Cao Renyi — Co-founder & Product Director, LibAI
- Cheng Weizhong (David) — Founder & CEO, Zhongke Shenzhi (中科深智)
2D video vs 3D path
Cheng: Contested. Many dismiss Yann LeCun as “outdated,” but his point stands: next-token / next-pixel prediction may miss that video is about representation space relations, not only pixel adjacency — world expressed via language, graphics, math, etc.
Zhang (Wujie): Agrees 3D+video stay important; today image is more mature for product control. Train industry models on design assets (e.g. fashion) with rich tags; enterprise data/standards → custom enterprise model. Case: Wanshili scarf LLM — designers + customers personalize patterns via flexible production.
Wujie stack: Multimodal platform (image primary; also video/3D/text); C-end app + Pro workflow mode; one-stop product/model/compute AIGC.
Why video (Jingying); multimodal (Zhizi)
Zhu: Scenario-first (TouchPal input / 10M DAU novel app background). GPT-3 showed GenAI content platforms — 2 years on video; v1 transformer ad gen → v2 diffusion short video. Still hard, but industry leapt in 6 months; expects major jump in ~1 year.
Gao: Building multimodal since 2020 (pre-CLIP/SD); vision–text alignment then SD/SVD-class video. Language needs near-perfect fluency; vision more forgiving → image gen feels “ahead.” Still needs transformer-class breakthrough for vision foundations. 3D/video add space/time → harder before image is mastered; new 3D/video startups still valuable atop academic base work.
Cheng: 3D often about motion/expression micro-problems; separate character model vs motion; future 2D character video may be driven by 3D motion/expression data.
Open source / build vs buy; commercialization
Cao (LibAI): 3D for precise control customers want; multimodal interaction (LLM metaphor understanding vs SD). Balance frontier R&D vs near-term commercial quality.
Zhu: Commercialization follows core product, not a universal open/closed dogma — own video models for short drama quality; buy mature TTS; stay open to third parties that strengthen the moat.
Further Q&A in CN on video monetization stage, LibAI global traffic #17 vs Wujie revenue, Zhongke Shenzhi commercial stage, tech vs product vs service, gen vs shoot share, creativity need, 2024 vision tech to watch, AI-native vs Big Tech — preserve full fidelity on ingest; EN keeps core technical/commercial lines above.
— END —