Unique Research · Insight / Unique Friend Q&A · January 12, 2024 · Guest: Zhang Yingfeng, Founder, InfiniFlow (英飞流)
Inventory type remains insight (Unique Friend series — not partner welcome).
Guest
Zhang Yingfeng — Serial founder: 7 years search engines; 5 years DB kernels; 10 years cloud/big-data infra; 10 years AI algorithms (ads, CV, NLP). Led three large enterprise digital transformations; supported C-end e-commerce with millions DAU and ~200M daily dynamic search requests.
InfiniFlow — Shanghai, Aug 2023; AI infra for universal enterprise AI benefit. Flagship: AI-native database for RAG with LLMs; full industry RAG solutions end-to-end.
Q&A
1. Product and core scenarios?
(1) Infinity AI-native DB for RAG — fixes vector-DB gaps serving RAG and future RAG evolution. (2) Full RAG solutions — DB alone is far from business; LLM landing is early; solutions accelerate digitalization.
2. R&D / GTM; biggest 2023 challenge and win?
DB open-sourced before end-2023 with community attention; first release ~Mar–Apr 2024 with seed trials. Parallel RAG solutions in multi-industry POC. Challenge: LLM hype vs enterprise-satisfying quality → hard paid conversion. Harvest: accelerating domestic RAG trials; customer feedback reshaped Infinity priorities toward present RAG pain points while keeping long view.
3. GenAI opportunities vs prior waves?
Digitization is no longer lip service. Prior AI stayed in limited marketing scenes; post-traffic-dividend needs efficiency/office apps — LLMs make general-purpose enterprise solutions possible (vs single-problem AI).
4. China landing challenges?
Domestic LLMs lag ChatGPT → scene limits; weaker pay willingness and expectation–reality gap; compute bottlenecks even with RAG — hope domestic GPUs/ecosystems catch up.
5. Competing with giants?
In infra, giants rarely win by dimensionality reduction — needs long accumulation; startups iterate faster with better usability. Giants often KPI wheel-reinvention, weak on features/usability/cost.
6. Are LLMs still too weak?
Far from ideal — severely limits apps. Benchmarks claiming near/over ChatGPT often collapse on real RAG scenes.
7. 2024 tech to watch?
Whether or not “Year of RAG,” fast RAG growth expected — plus RAG-boosting models (doc understanding/ranking). Multimodal + cheap open multimodal/MoE; inference accel via quantization/sparsity and new backbones. Hopes domestic/open LLMs break on real abilities (summarization, controllable hallucination) not just leaderboard wins.
1 CN image; attach on ingest. Signup CTA in CN omitted.
— END —