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

Summary

This 张小珺Jùn｜商业访谈录 episode interviews Luo Fuli / 罗福莉 about the shift from pretraining-centered chat products toward agent frameworks, post-training, long-context work, tool use, and model orchestration. Open Claw and Open Cloud are framed as a new middle layer connecting people, models, memory, tools, tasks, and cost routing rather than as ordinary UI shells. The episode also uses Memo VR, Xiaomi, and Luo’s organization practice to connect Agent Post-Training, Agent RL, Training Compute Allocation, Agent-Optimized Model Architecture, and AI Organization Design.

Key Claims

• Luo Fuli / 罗福莉 argues that the model race has entered a second act where Agent Harness design, Agent Post-Training, long context, tool feedback, and workflow data matter alongside base-model scale.
• Open Claw and Open Cloud are presented as modifiable agent frameworks whose memory, workflows, multi-agent structure, and skills can be adapted by users and teams.
• Agent frameworks can raise the ceiling of frontier models and also make smaller or mid-tier models useful in daily productivity tasks when the harness supplies context, tools, and routing.
• The source says post-training should shift from chat-oriented behavior toward agent-oriented work, including SFT/RL data generated through simulated user agents and multi-turn tool scenarios.
• Code is treated as unusually general training material because software work is long-horizon, multi-turn, dependency-heavy, verifiable, and close to real agent workflows.
• AI Skills are described as a way to encode organizational knowledge, execution norms, and private process details that are unlikely to appear in pretraining data.
• Model Harness Co-Evolution becomes the main competition thesis: strong models and strong agent frameworks adapt to each other, and the winner may be the team that iterates the loop fastest.
• Memo VR is described as a model series organized by role: Pro for understanding and complex scheduling, Omni for perception, and TTS for voice output, with Flash and Pro designed around long-context efficiency and speed.
• Agent-Optimized Model Architecture matters because hybrid attention, sliding windows, MTP, KV-cache tradeoffs, and architecture simplicity can leave room for post-training and agent adaptation.
• The source presents 1T-plus total parameter scale as an important entry ticket for the strongest agent-level competition, while still emphasizing architecture, post-training, evaluation, and cost.
• Training Compute Allocation changes because agents speed up idea generation and code execution; the bottleneck moves toward parallel experiments, evaluation, and enough cards for research and post-training.
• Agent RL needs infrastructure that spans agent frameworks, GPU/CPU/storage resources, rollout, fault tolerance, heterogeneous environments, and train-inference mismatch, not only a model inference engine.
• AI Organization Design is part of model capability: Luo favors flat, cross-stage collaboration without rigid pretraining/post-training groups or absolute leads, because creativity and fast migration across problems are treated as strategic.
• The source repeatedly links Research Taste to speed: when agents compress idea-to-evaluation loops from weeks toward hours or days, choosing the right ideas and validating failures matters more.

Key Quotes

“Agent 框架不是普通产品” — the episode’s distinction between UI and the thicker model-workflow middle layer.

“环境比经验更重要” — Luo’s hiring and team-development frame.

“卡的数量成为关键瓶颈” — the source’s compute-allocation claim for the agent era.

Connections

• Luo Fuli / 罗福莉 — guest and source of the model-training, organization, and agent-framework claims.
• Xiaomi, Memo VR, Alibaba, and DeepSeek — Luo’s current model-team context and prior institutional background as described in the episode.
• Open Claw, Open Cloud, Agent Harness, Agent Self-Evolution, and Model Harness Co-Evolution — agent-framework branch extended by the episode.
• Agent Post-Training, Agent RL, Training Compute Allocation, Agent-Optimized Model Architecture, and Frontier Model Scaling — model-training and infrastructure concepts added or extended by this source.
• AI Skills, Persistent Agent Memory, Long-Horizon AI, ML Coding, Multi-Agent Collaboration, and Model Workflow Fit — existing concepts the episode strengthens through agent-specific model training.
• Research Taste, AI Organization Design, and Problem Definition In Research — human and organization capabilities that become more important when agents accelerate experimentation.

Contradictions

• No direct contradiction with prior wiki content.
• The source is more optimistic about Open Claw and Open Cloud than the existing Probabilistic Software and Agent Permission Boundaries notes, but it does not remove those risks; it shifts emphasis toward model training, skills, evaluation, cost routing, and framework iteration.
• The source notes possible transcription or naming mixups around product names, so specific product timelines are recorded as episode claims rather than independently verified facts.