Unique Research · Interview · November 29, 2023 · Speaker: Wu Chenglin (吴承霖), Founder & CEO, DeepWisdom · From AWS × Unique Capital “AI Globalization Accelerator” Shanghai offline
Wu brings large-scale complex AI delivery experience at Tencent-class user/data scale; authored open multi-agent framework MetaGPT; world champion in NeurIPS AutoDL / AutoWSL / KDDCup Graph; papers in TPAMI / KDD / CVPR / ACL; Forbes 30U30, Hurun 30×30, and numerous Tencent/Huawei internal awards.
Keypoints: Agents add memory, planning, and tools on top of LLMs; Multi-Agents further fold in environment and SOPs. Core formula: code = SOP(team) — replace SOPs with an intelligent system and code generation can fully automate as SOPs and teams go online. Evaluate SOPs like code — with benchmarks. Boundary between what LLMs vs. Agents should do remains blurry. Agents excel at longer code planning/generation; the core metric vs. Copilot is leverage — Copilot’s is finite, Agents can make it unbounded. Natural-language programming is arriving fast. AutoGPT, AutoGen, and MetaGPT each have design strengths; MetaGPT’s research path is more “Android-like” — and hard.
1. MetaGPT in practice
Two thrusts: (1) a natural-language programming framework so users guide Agents to write code easily; (2) an Agent society emphasizing multi-agent collaboration. Late last year the team foresaw NL programming within ~two years — C → C++ → Java → Python ultimately returns to natural language; programming languages were designed for machines, not as humanity’s endpoint.
One NL sentence can map to unbounded lines of Python. Distinction vs. Copilot: Copilot leverage is capped; Agents unbounded. MetaGPT has 30,000+ GitHub stars; one-line prompts can generate small games at ~$1 per prompt line.
MetaGPT both constructs tasks and plays the user role — build something and use something as two sides of one coin — enabling self-iteration on complex games and CRM-class systems. In Minecraft, task-oriented strategies completed the diamond tech tree in 16 task rounds — 5× faster than VOYAGER. For Werewolf (multi-round group strategy), the team designed fine multi-agent communication topologies plus reflection and experience-learning mechanisms.
Tech ambition: push HumanEval Pass@1 toward 100. With MetaGPT they have reached an impressive 86 (GPT-4 reference ~67); data-based projection ~97 with low risk. When one-shot correct generation of every function nears 100, the core problem shifts from single-function generation to function-list generation + single-function generation — approaching unbounded code generation.
2. Code = SOP(team)
Example: build 2048 — multi-agent collaboration among product manager, architect, project manager, engineer, and QA: requirements, decomposition, competitor analysis, quadrant maps, demand pools — classic software-company SOPs simulating Scrum as a mini-waterfall.
Curious observation: ask for a Toutiao-like recommender architecture (not on GitHub) and MetaGPT draws a complete architecture diagram — acting as an information compressor that summarizes fragmented information into workable structure. Abstract every Fortune 500’s SOPs and you might simulate a new Fortune 500; put a software company’s SOPs online and you create an online software company. Formula: code = SOP(team).
3. Human-in-the-loop and self-improvement
Human-in-the-loop patches Agent weaknesses (e.g. vision-language gaps in text-centric GPT). One approach: a mid-tier human “hire” for UI drafts; MetaGPT also has an Agent UI via a fine-tuned Stable Diffusion model — not as good as humans, but workable. The framework can migrate beyond software cos — e-commerce CEOs in the community have moved business logic into the virtual world via code.
Stack: OS + AI Infra (train/fine-tune/infer/deploy, data/features) under MetaGPT as NL-level IDE; top layer AgentStore for Agent interactions and one-stop AI Infra so self-optimizing Agent groups accelerate model development and ops. Agents as “human-like structures”: OpenAI aims to scale GPT params ~100× toward brain scale, but completion ability alone doesn’t solve memory. Vector DBs may hit ~60–70% precision retrieving ~100 texts — not how brains work. Human memory has multiple systems (including procedural) and a critical forgetting mechanism — most things forgotten, important things retained as semantic, episodic, and procedural memory.
4. Human–Agent co-creation
Two concepts for multi-agents: SOP (encoding humans from practice, unified with machine encoding like Python in the digital world) and review. SOP data lives in head companies’ systems/docs and is hard to get — modeling needs close industry relationships. Review corrects hallucinations and can prevent ~95% of errors.
Human–machine collaboration evolved Embedded → Copilot → Agent → Society. We are in Agent mode. Future Agents might be company CEOs while humans are bot members with a kill switch.
5. Remake internet apps in the LLM era
Thirty years of information efficiency: yellow pages → search → recommendation; LLMs raise fuzzy-logic compute further. Paper reading can be LLM-digested then peer-reviewed at near-human quality — most internet apps deserve a remake. WeChat to Taobao are all reconsidering how to use LLMs — whether incumbents add the services or third parties take their traffic. Agents abstract brain logic; solve thinking and standardized memory and the Agent era truly arrives.
— END —