Unique Research · Interview · November 15, 2024
At Unique Capital’s AIGC Application Development Summit & AIGC100 Awards in Beijing (Oct 17, 2024) — 50+ speakers, ~1,000 practitioners — DeepWisdom founder & CEO Chenglin Wu delivered “Multi-Agents Leading a New Exploration of Code Generation.” Edited highlights:
LLMs + agents democratize natural-language programming
Software — whether frontier LLMs or classic NLP — faces four global pain points:
- Hard to code — many languages (Java, Python, C++…), DS&A, distributed systems for big data.
- Long cycles — requirements → design → code → test; each stage costly.
- Expensive talent — >27M programmers globally (→~28M in 2024); total hire cost >$2.79T (~$99k/person-year average).
- High barriers — IT outsourcing market ~$52.5B / outsourcing services ~$66.5B. GitHub Copilot can cut programming-related cost by up to ~90% via generation and routine automation.
Natural-language programming is getting funded: Poolside (ex-GitHub CTO) raised $500M seed at $3B valuation (eBay Ventures, Nvidia et al.). Copilot: 1M paid subs by Oct 2023; GitHub 2024 revenue expected ~$2B with Copilot >40% of growth.
Multi-agents lift LLM performance
Classic agent = observes & acts on environment — too thin; agents should exceed human mistakes (e.. autonomous driving). Analogy: LLM ≈ cognition (language/image/sound); cognition ≠ full reasoning — agents add observe / think / act / memory, trained in environments. Multi-agent systems add agents, environment, SOPs, review, routing, subscription, economic mechanisms.
MetaGPT (DeepWisdom, open-sourced Jun 2023): first multi-agent NL programming framework — 40K GitHub stars by Sep 2024; repeated #1 GitHub Trending. Data Interpreter drew Andrew Ng’s attention. Automates docs and roles (PM → architect → eng → QA) via SOPs + metaprogramming to fight LLM hallucination on project-scale codegen. Internal experiments (Python games, CRUD, simple analytics): average score 3.9, beating peer auto tools.
Representative cases
Android Agent: learning framework so agents acquire tap/swipe/long-press/double-tap semantics across icons without prior phone experience.
MetaGPT X (academic): multi-agent software problem-solving — 46.67% solve rate with advanced code understanding/debug. Open-sourced AFLOW: MCTS auto-optimizes agentic workflows (nodes/edges + operators) — +5.7% vs hand design; better performance at 4.55% of GPT-4o cost. Workflows as “thought guidance” can yield o1-like effects.
Data Interpreter: with university partners — hierarchical graph dynamic planning, tool integration/generation, confidence + experience-driven reasoning. SOTA: MATH +26% accuracy; ML-Benchmark 0.86→0.95; open-ended tasks +112%. Experience reuse cuts cost after first solve.
SELA auto-experiment system: LLM + MCTS for AutoML — on 20 datasets, win rate 65–80% above prior methods.
Closing
Ultimate multi-agent form? Org design matters (Tencent/Alibaba/Huawei vs Amazon/Microsoft/Google; Apple’s centralization also works). Thousands of companies build agents — will they form networks like humans on WeChat? Wu: agent nets inside every company, phone, car — different functions that compose into an ecosystem.
— END —