Podcast Atlas

Vol. 165 做客声东击西：「龙虾」和 vibe coding 正如何改变我们的思维

source Updated 2026-07-07 Tags: Podcast, Ai, Agents, Vibe-Coding, Work

Summary

This 枫言枫语 crossover with 声东击西 brings Justin Yan, 徐涛, and 王俊玉 together to discuss Open Claw and Vibe Coding from engineer, non-technical power-user, and entrepreneur/product perspectives. The episode uses 声动活泼’s internal AI Hackathons and Xu Tao’s own news-crawling and topic-recommendation prototypes to show that AI coding lowers the threshold for turning workflow pain into software. Its main contribution is a work-and-training frame: Agent Native Software, AI Skills, Persistent Agent Memory, and proactive loops can make agents feel like trainable digital colleagues, but production systems, permission boundaries, entry-level training, and high-end human judgment still matter.

Key Claims

  • 声动活泼’s AI Hackathon showed non-technical employees building audio, title-brainstorming, and image-related tools for real work pain, turning Vibe Coding from an engineer tool into a broader organizational capability.
  • 徐涛’s “small white user” experience is that the shock of Open Claw is not only chat quality, but the discovery that an agent can solve tasks by writing programs, using back-end logic, fetching news, and pushing results.
  • Justin Yan frames Open Claw as an early historical turn toward Agent Native Software: software can become self-changing and agent-centered rather than fixed button-result software.
  • 王俊玉 emphasizes three OpenClaw lessons: proactivity, long memory, and AI Skills, with skills acting as reusable work methods rather than only user-preference memory.
  • In company settings, OpenClaw-like agents can observe Slack, Linear, GitHub, Feishu, or similar systems and wake on a schedule, but that makes Agent Permission Boundaries and responsibility design unavoidable.
  • Vibe Coding is strong for clarifying demand and creating prototypes, while the jump from prototype to stable company system still needs architecture, debugging, reliability, and AI Coding Verification.
  • The episode treats many “sense,” taste, editorial, and product-design judgments as partially decomposable into criteria, making AI useful for surfacing tacit standards without fully replacing expert judgment.
  • AI performs best where outputs are quantifiable, repeatable, and process-driven; low-data, highly contextual, relational, and taste-heavy work remains more dependent on Human Judgment Under AI.
  • Entry-level work and professional training may be disrupted because AI can absorb some beginner tasks, but students and young workers can also build richer projects earlier if they practice with the tools.
  • Managing AI is framed as a management skill: setting goals, processes, review points, and escalation rules can make one person act more like a front-line manager of multiple agents.
  • The guests’ advice is practical rather than purely optimistic: use AI at meaningful intensity, try building small things, keep foundations strong enough to judge output, and move toward uniquely human strengths such as taste, creativity, and understanding people.

Key Quotes

“小白、创业者和工程师” — the episode’s framing of three perspectives on OpenClaw and vibe coding.

“可量化、可流程化、可重复” — the episode’s boundary for where AI tends to work best.

“悲观者正确，乐观者成功” — the closing attitude toward AI disruption and personal action.

Connections

Contradictions