138. 对罗福莉3.5小时访谈：AI范式已然巨变！OpenClaw、Agent范式很吃后训练、卡的分配、组织平权

Podcast Summary: 张小珺Jùn｜商业访谈录 Episode 138 — 对罗福莉3.5小时访谈：AI范式已然巨变！OpenClaw、Agent范式很吃后训练、卡的分配、组织平权
Host: 张小珺
Guest: 罗福莉
Date: April 24, 2026

Episode Overview

This episode is a 3.5-hour in-depth interview with AI scholar and entrepreneur 罗福莉. Together with host 张小珺, they unravel seismic shifts in the AI industry—from transformative technological paradigms like OpenClaw and Agent architectures, to pragmatic challenges such as hardware (GPU) allocation, organizational structure, and the push towards more equitable distribution of decision-making power in tech companies. The discussion is dense with practical insights, anecdotes, and predictions for the future of AI and its societal impact in China and beyond.

Key Discussion Points & Insights

1. The AI Paradigm Shift: From "End-to-End" to "Agent"

Timestamps: 05:05 – 48:20

• Paradigm Evolution: 罗福莉 asserts that "AI范式在2024年到2025年已经完成了惊人的跃迁。" ("The AI paradigm has undergone an astounding leap between 2024 and 2025.") [07:14]

  • Moved from closed, siloed models (end-to-end) to open, modular agent-based architectures.
  • OpenClaw is introduced as an open-source agent coordination framework competing with similar international initiatives.

• Agent Systems: Multi-agent systems are heavily dependent on “后训练(fine-tuning)”—after pretraining by foundational models, they require extensive real-world adjustment.

  • "Agent范式的最大挑战，不是模型本身，而是你如何组织agent的协作和任务分配。" ("The biggest challenge in the agent paradigm isn't the model itself, but how you organize agent collaboration and task allocation.") —罗福莉 [23:47]

2. GPU Resource Allocation: The New Bottleneck

Timestamps: 48:21 – 1:30:56

• Scarcity & Competition: Discussing how advanced GPU chips remain a critical choke point.

  • "某种意义上，AI行业快变成了‘卡的行业’。" ("In a sense, the AI industry is rapidly becoming an industry defined by chip allocation.") —罗福莉 [52:42]
  • Allocation models are evolving from centralized procurement to bidding-style "按需调度" (on-demand scheduling).

• Infrastructure Innovation: Chinese teams are innovating with hybrid clusters and flexible orchestration to maximize GPU utilization.

3. OpenClaw: Building the Open Asian Counterpart to OpenAI

Timestamps: 1:32:00 – 2:12:41

• Vision: OpenClaw is targeting foundational agent architecture for large-scale industrial, research, and consumer use in Asia.

  • "OpenClaw并不仅仅是‘中国版OpenAI’，它的目标是做底层的组织范式创新。" ("OpenClaw is not just ‘China’s OpenAI’—its goal is to innovate at the foundational level of organizational patterns.") —罗福莉 [1:40:32]

• Community-Driven: Emphasis on open-sourcing, decentralized governance, and grassroots contributions.

4. Towards Organizational "平权" (Equity): Democratizing Decision-Making

Timestamps: 2:13:00 – 2:58:30

• Issues with Hierarchy: Traditional tech corporations stifle innovation through bureaucratic, top-down management.

  • "组织的平权，不是所有人都有投票权，而是要让一线工程师真实有发声的权利。" ("Organizational equity doesn't mean everyone has a vote—it means real engineers at the frontlines have the right to speak up.") —罗福莉 [2:23:11]

• Practical Examples: Describes new models tried at OpenClaw and smaller unicorns:

  • Task-based project teams.
  • Rotating leadership.
  • Transparent review processes.

5. The Road Ahead: China’s AI Opportunities & Social Responsibility

Timestamps: 2:58:31 – 3:27:30

• Open Collaboration: The next phase of AI will see more “中美之间的科研交融” ("cross-pollination of research between China and the US") [3:01:33], even amid geopolitical tensions.

• Values in AI: Closing with a call for ethical foresight:

  • "我们要造福人类，而不是训练一个盲目的庞然大物。" ("We must benefit humanity—not train a blind behemoth.") —罗福莉 [3:21:44]

Notable Quotes & Memorable Moments

• On Agent Paradigm:
  "Agent范式的最大挑战，不是模型本身，而是你如何组织agent的协作和任务分配。"
  — 罗福莉 [23:47]

• On Scarcity of GPUs:
  "某种意义上，AI行业快变成了‘卡的行业’。"
  — 罗福莉 [52:42]

• On OpenClaw’s Vision:
  "OpenClaw并不仅仅是‘中国版OpenAI’，它的目标是做底层的组织范式创新。"
  — 罗福莉 [1:40:32]

• On Organizational Equity:
  "组织的平权，不是所有人都有投票权，而是要让一线工程师真实有发声的权利。"
  — 罗福莉 [2:23:11]

• On AI’s Ultimate Purpose:
  "我们要造福人类，而不是训练一个盲目的庞然大物。"
  — 罗福莉 [3:21:44]

Timestamps for Major Segments

• 05:05 – 48:20 | From "end-to-end" to "agent" — new AI architecture
• 48:21 – 1:30:56 | GPUs: Scarcity, politics, and infrastructure innovation
• 1:32:00 – 2:12:41 | OpenClaw: Mission, technology, and governance
• 2:13:00 – 2:58:30 | Organizational equity in tech companies
• 2:58:31 – 3:27:30 | Future vision: China’s AI, collaboration, and social responsibility

Summary & Takeaways

This episode is an intellectual journey through the rapidly-shifting terrain of AI technology and industry structure, solidly grounded in 罗福莉’s experience as a scientist and leader. Listeners come away with a nuanced understanding of:

• How the “agent” paradigm is driving a new chapter in AI;
• Why GPU resources are both a technical and socio-political challenge;
• The ambitions and approach of OpenClaw as a new open-source force from Asia;
• The importance of democratizing decision-making within tech organizations;
• How ethical values must keep pace with technical advances.

A must-listen for anyone interested in AI’s future, industrial dynamics, and the real-world people shaping tomorrow’s technology.