Unique Research · Insight / Unique Friend Q&A · October 12, 2024 · Guest: CEO, Lanma Technology (澜码科技)
Lanma was earlier welcomed as Unique Partner (EN package 20240903a066ij00); this piece is the Unique Friend Q&A — inventory type remains insight.
Product: AskXBOT — self-built enterprise AI Agent platform (design, develop, use, manage, knowledge). Conversational Agent creation for due-diligence review, sales assist, etc. Live in finance, energy, government, manufacturing, retail, education.
Moat thesis: Expert knowledge sets the Agent ceiling. LLMs lack firm-specific knowledge (star graduate without practice) and blur capability boundaries → hallucination. AskXBOT deposits SOPs, processes, rules, facts into a knowledge base; Agents help govern knowledge; expert knowledge powers Agents/workflows — a closed loop.
GenAI / pricing: AI-native on LLM extract / instruction-follow / reason. Fees today ≈ share of customer’s AI compute spend. Future: price like digital employees replacing junior/mid labor (e.g. ¥8,000 junior salary → ¥5,000 digital peer).
Tech lead: Absorb successive model capability jumps; scarce internal (not public) data — expert experience described in NL — as the flywheel via short/long-term Agent memory.
Ecosystem: Start from the desired future stack (infra / platform / app, like early internet). Lanma = platform layer; map model vendors and app vendors, then ally.
Challenges / China vs SV: Fast PMF chase under changing tech + acceptance → stay org-agile. Scaling-law view: SV clusters already ~100k GPUs; China has plentiful 10k clusters but not yet 30k+ unified clusters → new architectures/emergent abilities likely appear in SV first. China apps/scenes have data edge; multimodal may ride short-video strength; on-device (AIPC/AI phone) lag only ~3–6 months; core reasoning / O-series-class innovation gap larger and may widen with talent/compute/data.
2024 watch: How fast big SV labs follow O-series; multimodal breakthroughs that change input modes and unlock demand.
— END —