Unique Research · Insight / AIGC Online Open Day · September 27, 2023 · Full editorial rewrite of ~3,896 CJK — four company intros + full Q&A preserved
Unique Open Day on AI Agents — advantages, enterprise landing, and coexistence with humans. Guests: Lanma Tech, Waveform Intelligence, CAMEL-AI.org, AutoAgents.ai.
Companies
Lanma Technology (澜码科技)
Founded 2023; vision: “everyone can design their own AI Agents.” LLM-based next-gen automation platform — a “data flywheel” that learns skills from PC/phone/meeting scenes to build proprietary multimodal foundation models.
Waveform Intelligence (波形智能)
AI interactive-content platform from tools into full entertainment-chain coverage. Own creation tools: rich content, interactive, personalized, multimodal — beyond wrapping GPT APIs. Differentiator: interactive long-text generation with infinite interactive content; vertical interactive models via full-stack NLP LLM capability. Paper: RecurrentGPT on the initial framework.
CAMEL-AI.org
Open-source community for autonomous & communicative agents — study behavior, capabilities, risks at scale; ships agents, tasks, prompts, models, datasets, sim environments. Goal: understand human–AI coexistence — cooperation/competition forms, new economic/legal/education systems, accelerate science/engineering.
AutoAgents.ai (未来式智能)
China AI/cloud firm on the front line of enterprise digitalization. Team from Alibaba DAMO, Tencent, ByteDance, AWS, Google — LLMs, multimodal, self-evolving agents, intelligent decision. Product: agents that plan/decompose from user goals, connect data/apps. Scenes: intelligent CS, customer-centric sales, employee-centric recruiting, business analytics.
Roundtable
Trends next few years
Zhou Jian @ Lanma: Breakthrough moment — LLMs regenerate software, especially language understanding; “super individuals” will wield LLM skills for many tasks.
Jiang Yuchen @ Waveform: Agent power tracks base LM power. Email-class tasks: base LMs already suffice; agents mainly provoke the model. Literary creation: base LMs fall short — agents matter more. Agents become standard in every app; choose train-own / train-with-framework / API by scene.
Li Guohao @ CAMEL: Near term, high-accuracy scenes still need human-in-the-loop (often passive collaboration) — science, software. Entertainment/creative (novels, game AI, animation) easier (imagination > absolute accuracy). Academic priority: controllable generation under format constraints.
Yang Jinsong @ AutoAgents: Uneven landing as models/tools iterate — education/CS first; high-stakes later. Long term: multiple intelligent services around each user.
Advantages, problems, impact
Lanma: Hardest: expert tacit knowledge (e.g. spotting résumé fraud needs context). Agents that capture tacit/operational/embodied knowledge differentiate from public LLMs; enterprises keep that knowledge private — large commercial opportunity.
Waveform: Is an “agent” only a GPT API, or layers beyond LLM (e.g. ontology)? Content creation needs both model and agent-framework innovation plus business-model innovation. Entrances change; basic human needs don’t — tech must serve hard needs.
CAMEL: Role split: humans higher-level; agents call APIs for concrete decisions — risk of unsafe/biased outcomes (e.g. hiring). Alignment is hard to define; safety is clearer — do not harm humans.
AutoAgents: Agents disrupt SaaS/ToB UX — intent understanding, creative scene-adaptive interaction. ToB software as middle layer to data tables may shrink as users talk to tables via agents. Generated software cost becomes measurable → usage-based pricing vs feature bundles. Deep domain knowledge remains a moat; more companies sell APIs than monolithic SaaS.
Enterprise landing blockers
Lanma: Compute, data management, integration, privacy compliance, informatization maturity.
Waveform: Non-standardized enterprise data; fuzzy requirements/acceptance — users expect ~100% accuracy models rarely deliver.
CAMEL: Broader society questions — coexistence, cooperation/competition, ethics/law as agents evolve.
AutoAgents: Weak reasoning/planning and hallucination — narrow to scenes; constrain LLMs; three-layer best practice (industry / enterprise / individual) to raise completion probability.
Coexistence
Lanma: Agents may approach GAIA-like sensing/control of society — ethics, self-awareness, motives; steer for long-term human interest.
Waveform: More social bonds with virtual agents in entertainment; productivity + spiritual richness, with interpersonal-relationship concerns.
AutoAgents: Short-term productivity vs resource imbalance and self-iterating agents that drift from human interest — ethics, regulation, global collaboration.