Unique Research · Unique Viewpoint · March 26, 2024
Background
On January 24, 2024, Unique Capital and partners held the 2024 AIGC Application Development Summit in Shanghai, with nearly 1,000 AI practitioners attending. This piece is edited from the guest roundtable segment.
Guests
- Lu Yilei — Founder & CEO, Baihai Tech (白海科技)
- Zhang Yingfeng — Founder, Infinity (英飞流)
- Qin Xuye — CEO, FutureSpeed (未来速度)
- Lin Qunshu — Founder & CEO, Integer Intelligence (整数智能)
- Host: Li Jiarui — Founder, Juzi.bot (句子互动)
Q: Looking at the LLM tech stack over the past year, what stood out most?
Lin Qunshu (Integer): Before Transformers, datasets were smaller and hand labeling was enough. With large text and multimodal models, hand labeling collapses. Teams moved to automation — iterating algorithms on previously labeled data — and into domains like intelligent driving (AD + cabin) that need huge high-quality datasets. Synthetic data is rising where real collection is hard (e.g., AV). Enterprises that once skipped data-engineering platforms are now buying automated labeling, cleaning, and review.
Qin Xuye (FutureSpeed): ChatGPT may be strongest overall, but enterprise adoption hinges on (1) security — finance clients often cannot use certain APIs; (2) scenario fit, not absolute SOTA; (3) cost — if a smaller fine-tuned model works, don’t upsize. Our platform treats models as interchangeable with one-click switch.
Zhang Yingfeng (Infinity): Open source pushed the whole stack — models, inference, middleware — toward cheaper enterprise readiness. Expectations for LLMs started high in early 2023 and have cooled as the stack industrializes from C toward B. On infra, SJTU’s open-source PowerInfer (ordinary GPUs for large-model inference) is rare domestic originality, even if not yet production-ready.
Lu Yilei (Baihai): Stable Diffusion (Oct 2022) was interesting; by spring 2023 LLMs overturned software architecture — from distributed per-scenario AI to foundation-model-centric integrated AI, raising compute and data bars. We burned ~¥3M compute building Chinese models on open bases with poor early results; now Chinese understanding in domestic models is strong — underscoring data. For domain models we insist: better 10k high-quality samples than 1M noisy ones.
Q: What is Baihai’s distinctive edge?
Lu: Baihai is an AI train/infer cloud platform for fine-tuning and full-lifecycle traditional AI — with an open-sourced AI IDE in Rust, strong GPU scheduling, and usability so developers, analysts, even business users can build models. Auto + human eval for accuracy/relevance/consistency; strict review before feedback data reflows (to avoid GPT-4-style quality drift from unclean feedback). 2023 contract value reached tens of millions RMB; dozens of Fortune 500 / SOE clients. Positioning: middleware for the last mile of LLM apps — partner, don’t compete with foundation-model/BAT compute providers.
Q: Infinity’s AI-native database and RAG — what’s new?
Zhang: RAG cannot be vector-DB-only. Strong RAG often comes from search companies (e.g., Baichuan/ex-Sogou; overseas open-source Vesper). Two requirements:
- Infra: multi-path recall — vectors + search (precision) + structured query. The store must first be a real database (tables, ACL, accounts). Infinity’s “AI-native DB” thesis: one enterprise LLM digitization layer + one DB that holds all enterprise data for the model.
- Data: end-to-end open-source RAG including prep / ETL-for-AI to organize data for better retrieval. Fit domains: finance, IT, industry, healthcare, law — knowledge-heavy internal corpora. Industrial buyers are cost-sensitive (tickets may be thousands–tens of thousands RMB vs finance at hundreds of thousands+).
AI-native DB / RAG for the first time unlocks infra’s path from B2B toward consumers — everyone may eventually have a personal RAG / cheap personal knowledge base (e.g., uploading meeting audio). That needs cheaper models and cheaper DB infra.
Q: FutureSpeed’s Xinference — tech edge and commercialization?
Qin: Open + commercial editions; GitHub open edition ~2,000 stars. Solves model variance, sizes, load/throughput; OpenAI-compatible REST API; integrates LangChain, LlamaIndex, Dify.AI; engines Transformers / vLLM / GGML; multi-node distributed deploy; domestic + foreign GPUs including Mac GPU. Commercial edition targets enterprise performance, observability, HA — finance/government with less API dependence, and cost-sensitive verticals like customer service. Ecosystem partners (Infinity, Dify, etc.) matter: China buyers want integrated solutions, not point products.
Q: Integer’s one-stop data management — core advantages?
Lin: Two words: automation and compounding. Founded from quant/NLP labeling pain — domestic vendors were either AI-naive or pure hand-label shops. Automate 80–90% of labeling; humans review. Algorithms must improve over time (MLOps in-platform) so PMs without deep algo skills can retrain pre-labelers — driving data cost down so previously blocked AI scenarios become viable.
Q: How does Integer stay ahead in a crowded AI data market?
Lin: Intelligent driving is extreme: OEMs may process >100M km/year of camera + LiDAR. Tesla’s early hand labeling could take >500 hours per video clip; their closed-loop + 4D tooling lifted efficiency ~100× over ~3 years. In 2023 Integer shipped a 4D labeling platform — fuse multi-frame via 3D reconstruction so one annotation covers the sequence. Vertical customization for OEMs; claimed lead vs competitors (tools ready Mar–Apr 2023 vs peers mid-2024). Transfer overseas AI data-engineering practice to cut Chinese AI companies’ data costs.
Q: Where is the next tech explosion in the LLM stack?
- Lin: 2024 watch synthetic data — beyond simple augment. Integer will launch an open data foundation with ¥10.24M (2^10) to create frontier engineering datasets.
- Qin: 2024 as the year of real landing — models, agents, and full-stack readiness crossing thresholds.
- Zhang: Landing year; focus on RAG — if RAG fails, agents fail; bottleneck still foundation models, with expected breakthroughs this year.
- Lu: 2023 = foundation-model year; 2024 = tooling year for startups (train/infer, agents, RAG platforms) — without tools, no fast landing.
2 CN images; attach on ingest.
— END —