Unique Research · Recap · June 6, 2024
Unique Research distilled this Open Day on how B2B industries co-evolve with foundation models and AI — speaker highlights plus roundtable Q&A. Photos attach on ingest (11 images).
Cheng Junjie — Founder & CEO, Chuansheng Marketing (传声营销)
1. Data is the core asset; every activity must orbit data generation, analysis, and application. In the AI era, knowledge is the new high ground — knowledge bases fuel intelligence — but the durable lever for productivity remains workflow innovation.
2. Users want workable solutions, not magic tools. Copy tools must rest on solid copy theory; first define great structure, then use AI.
3. PUJD model for copy: Pain Point → Unique solution → Joy Point → supporting Data (e.g., 30% energy save; ¥30k investment → ¥3M savings). Structured input is distilled from excellent copy patterns. For intelligent CS, jumping straight to Agents can be imprecise — product description theory must come first. Evolution from 3W1H (Why / What / How) to 3W2H (+ Who / How Much) — case knowledge + pricing/commercial terms — plus unstructured non-product knowledge.
Lu Yue — VP, Yida Tech (壹沓科技)
1. Four stages: standardization → informatization (human→software) → digitization (software→human, e.g. RPA; ≥50 definitions of “数字化” in their survey) → digital-intelligence (software→software→Agent; Gartner hyperautomation: task → process → cognitive → hyperautomation with LLMs).
2. Almost every industry enters “AI+.” Narrow AI vs AGI (reason, plan, learn, communicate, perceive). Rebuild commercial logic.
3. Three work modes: Embedding → Copilot (chef + prep robot) → Agent (instruct; agent executes).
4. Today’s UIs are for humans (freeze workflow → role UI). Tomorrow Agents may hit databases directly — no classic UI. Commerce entry may be a dialog or humanoid: search, compare, buy on preference.
Liu Yuchen — Founder & CEO, Yiyan Tech (意言科技)
1. AI Agents decompose tasks and self-reflect — System-1 execution plus System-2 analysis; learn and adapt end-to-end like humans.
2. Agents chase goals with environment feedback and tools; LLMs unlock strategies that pre-LLM agents lacked.
3. Scale alone isn’t enough — add reasoning, reflection, and tool-calling loops.
4. Andrew Ng: Agent workflows are undervalued; GPT-3.5 + agent loop can beat bare GPT-4. Four patterns: Reflection, Tool Use, Planning, Multi-agent Collaboration.
5. Fast win — sales: mine historical calls/notes for intent and visit progress; low-risk, wide coverage.
Roundtable Q&A (abridged, full fidelity to CN)
Q: How much do base LLM capabilities matter?
A: Raw LLMs hallucinate (e.g., “681 × 2526?”). Precise apps use function calling / tools (Python, etc.). Depth of tooling depends on business complexity — hundreds of nodes vs a simple search/companion bot.
Q: Can AI help B2B marketing collect competitor intel?
A: Classic social listening — once costly (servers + heavy custom). LLMs now find, analyze relevance, and draft reports cheaper.
Q: Who buys Agents in B2B? Where to land first?
A: Decision with the department that benefits — CX/marketing for service bots; sales VP for pre-sales bots; ops/production for training Agents. Whoever benefits pays.
Q: Will GPT-6 replace Agents?
Liu: Transformer LLMs compress world knowledge — strong at summary, weak at true creativity/reasoning unless new paradigms appear; more params can look like reasoning without being it.
Lu: Base-model progress alone won’t overturn verticals — Altman noted even GPT-5-class models struggle where private enterprise data matters; China focus on private vertical models + Agents.
Cheng: Private knowledge and workflow sediment stay stable — irreplaceable IP regardless of model generations.
Q: LLM value in CRM?
A: (1) Faster/cheaper existing workflows (reports that once cost ¥100–150k). (2) Agent-to-agent and workflow chaining across SaaS A/B/C — CRM notes → Excel → report → email — rebuilt as connected Agents.
Q: Platform vs vertical? Agent vs RPA? NLP vs LLM coexistence?
A: RPA simulates mouse/keyboard; Agents driving RPA is a key path — complementary, not contradictory. Don’t treat LLMs as universal — big data mining and RPA still win some jobs. In CS, classic NLP intent/keywords still needed for latency. Fine-tuned small models can score/classify with less hallucination; mix by scenario.
Photos / slides attach on ingest.