Unique Research · Industry insight / roundtable · March 20, 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 roundtable. Note: Lanma Technology (澜码) founder Zhou Jian appears as a panel guest — not a partner welcome retag.
Guests
- Zhou Jian — Founder & CEO, Lanma Technology (澜码科技)
- Yang Jinsong — Founder & CEO, Future-Form Intelligence / AutoAgents.ai (未来式智能)
- Liu Yuchen — Founder, Tyrion AI
- Wang Shouyan — Founder, Deep Insight (DeepBI / DBI)
Q: What are you shipping in LLMs / agents?
Wang (Deep Insight): Product DBI — AI-driven data insight, not classic BI. Traditional BI assumes humans analyze and prioritizes visualization; DBI assumes AI should autonomously analyze raw data and surface insights (“data-led insight”).
Liu (Tyrion): AI-agent development platform + end-to-end agent solutions — expert knowledge bases, expert agents, and process automation for agent-doable tasks.
Yang (Future-Form): Multi-agent solutions for global enterprise scenarios — orchestration for knowledge-era AI automation pipelines.
Zhou (Lanma): Started before “agent” was trendy — LLM automation for business processes. From RPA days: workers handle data, documents, apps, processes → RAG for docs, text-to-API for data. Agents must be process-aware (know which stage) to prioritize work — key difference vs copilots.
Q: Brand scenarios — data, users, traffic; how do agents reshape them? (Liu’s brand-side background)
Liu: Early ChatGPT experiments analyzing BA ↔ private-domain consumer chat logs beat classic NLP on semantics/attitude/entity mention. Expert knowledge once lived in salespeople’s notebooks — digitize for training and offline→online handoff. Optimistic on foundation models + middle-layer agents/RAG as B-side infrastructure improves.
Q: Lanma AskXBOT’s three steps → ToB landing?
Zhou: (1) Digitize expert knowledge — knowledge leaves with experts (consultancies as lead magnets). Agents can ingest Excel dumps from unintegrated procurement/finance systems and write API multi-dim tables — CEO gives an account, not a system rewrite. (2) CUI flexible interaction — leaders won’t learn 300-field wide tables; build semantic + metric layers so machines adapt to humans. (3) Domain knowledge loop — JD/resume screening example: humans correct machine scores; rules for “job stability,” training-mill resumes, etc. become digital; experts scale from 100→1000 cases with AI assist.
Q: AutoAgents.ai “async collaborative” agent orgs?
Yang: Serial multi-agent patterns (AutoGPT / Stanford Town) amplify hallucination in enterprise use. Real work is parallel. Design: central scheduler agent decomposes tasks; sub-agents work in parallel with mutual cross-checks → usable outputs. Short-term: hallucination control; long-term: compresses human knowledge collaboration — LLM as brain, agents as roles/teams.
Q: Beyond query — DeepBI’s deeper value?
Wang: Visualization is for humans; models are token-sensitive to raw data. Goal: thinking → deep insight via collaboration + agent organization. Products: assisted analysis (any size table, ask anything) and auto analysis (machine explores without looping) — “infinite thinking” BI via agents.
Q: Third-party LLMs as black boxes — control & vendor relations?
Wang: Models ≈ elite universities graduating en masse — still hallucinate; orgs need training, constraints, workflows. Value models for domain knowledge + logic; occasional errors can spark exploration.
Liu: Models = gasoline; apps = cars; middle layer = engine/transmission — workflow, RAG, prompt, sometimes fine-tune + multi-model routing. Like cloud providers vs industry software.
Yang: Models = bricks; scarce skill is scenario landing. App side feeds model vendors (longer context, stable JSON, eval which model wins which task).
Zhou: Three value tiers: (1) integrate formerly unusable docs/chats via RAG — open 6B/13B often enough; (2) data analysis — medium difficulty; (3) master agent that actively asks / GPT-4-class function calling — likely 100B–trillion-param, costly. Apps are the “menu”; don’t use premium meat for mapo tofu. Moat: private SOP + private data enterprises won’t put on public SaaS — “palace kung pao with ordinary ingredients.”
Q: AI-native vendors — new customer relationships?
Wang: Classic Chinese ToB: ¥100Ms revenue needing 1,000+ people because rigid codebases. Math-trained colleagues now adapt across industries — tens of engineers for 10 verticals vs 100–200 before → ToB cost collapse.
Liu: Agents let customers edit behavior in visual UIs without vendor redeploys; future NL goals + history learning → business users configure agents. Outcome-oriented SaaS.
Yang: Buy results not tools (vs 2-week SAP training). Pricing by completed replies/tasks. Delivery shifts from algo experts to IT/business users — text-to-agent shortens long enterprise delivery cycles.
Zhou: Outcome-share pricing vs SaaS seat pricing. Example: ~¥1 cost for high-quality resume recommendations lifting interview adoption 2–3×; Softcom Power (~80k staff / ~2k recruiters at ¥6–8k salaries) would pay ¥5–10 per better shortlist — digital-employee economics from RPA days becoming real.
Q: Advice for AI founders in the new year?
Wang: ToB can become decreasing marginal cost. Entering existing markets (e.g., $10Bs BI) needs ~10× better product — AI now makes that possible at ~1/10 cost.
Liu: 2023 was literacy; by year-end brands planned real AI projects (phase 2–3). Optimistic on 2024.
Yang: Human–AI hybrid workflows (e.g., 70–80 of 100 nodes by AI); agents as new entry alongside web/app. Incremental scenario value.
Zhou: Agent count doubling ~every 12 months; inference cost /10 every 18 months; training /4. Zhipu GLM-4 ~3–6 months ahead of his GPT-4 catch-up expectation; domestic models to catch GPT-4 this year. CUI consensus exists but 2023 lacked scenarios due to hallucination/cost/learning — opportunity for agent cos. Zhipu valuation ¥2B→¥20B creates revenue pressure; 150 of China Top 500 already partners — need real business landing = agent vendors’ upside.
4 CN images; attach on ingest. Lanma kept as interview guest (inventory type interview), not partner.
— END —