Podcast Atlas
source Updated 2026-07-08 Tags: Podcast, Ai, Frontier-Models, Ai-Coding, Agents

136. 全球大模型季报第9集：和广密聊，Coding是AGI第二幕、硅谷御三家真相、模型正成为新一代OS

Summary

This 张小珺Jùn｜商业访谈录 episode argues that AI has moved from a chatbot-centered first act into a coding-agent second act, where Claude Code, Codex, and related tools can complete valuable work and accelerate AI research itself. The strongest synthesis is AGI Three Acts: chatbots create broad access, coding agents automate a large share of digital knowledge work, and automated AI researchers become the third act if models can run experiments, analyze results, and improve future systems. The episode also adds Model As Operating System, treating frontier models from Anthropic, OpenAI, Google, Meta, and xAI as possible platform infrastructure for future work, life, research, and agent ecosystems.

Key Claims

  • Coding is framed as the new AI accelerator: if agents can write, test, review, and iterate code, they shorten the experiment loop for software, data processing, multimodal research, and model development.
  • The guest divides AGI progress into three acts: ChatGPT-style chatbot access, Claude Code/Codex-style coding agents, and automated AI researchers that can help with science and model progress.
  • Anthropic is portrayed as the clearest short-term winner in coding because it focused resources, data work, product form, and organization around Claude Code instead of spreading attention across too many consumer and multimodal bets.
  • OpenAI is described as still strong and possibly capable of overtaking because its exploratory culture may discover new paradigms, but its ChatGPT consumer success is treated as a strategic distraction from coding.
  • Google and Gemini are described as a stable long-term competitor because of compute, cash flow, TPU, operating systems, and Workspace distribution, even if Google is said to have underweighted coding for several months.
  • Meta is framed as the most plausible fourth challenger, especially if it can integrate Manus or other product layers; xAI is described as strategically less focused but still resource-rich because of Elon Musk’s infrastructure and adjustment capacity.
  • The source treats Agent Harness as the work-environment layer for agents: models need tools, computers, accounts, permissions, task environments, and feedback loops rather than only chat windows.
  • Domestic model companies such as Kimi, MiniMax, Zhipu AI, and Doubao are described as moving toward the Anthropic-style high-value-task route, though Doubao remains the strongest domestic consumer-assistant example.
  • AI Investment Metrics should shift from DAU and consumer time spent toward Token Usage, high-value task completion, and whether model usage creates profit, savings, or new capability.
  • Token Maxxing is extended by the claim that a small number of high-value coding and agent users may produce more revenue than very large numbers of light consumer subscribers.
  • Model As Operating System is the episode’s platform thesis: leading models could become more important technical infrastructure than current search, mobile OS, or app platforms because applications and agents may sit on top of them.
  • The social-risk section intensifies Intelligence Devaluation and Human Resource Deflation Compute Infrastructure Inflation: knowledge, software, consulting, outsourcing, and junior white-collar work may become cheaper faster than institutions can absorb.
  • The personal advice is not simply “learn tools”; it emphasizes creativity, taste, judgment, and active AI adoption as the scarce layer once generic cognitive execution gets cheaper.
  • The investment thesis concentrates on companies that can continuously produce SOTA models, while treating robots, AI For Science, agent infrastructure, and One-Person Company experiments as important but secondary portfolio branches.

Key Quotes

“Coding 是新的 AI 加速器” — the episode’s core claim that coding agents accelerate AGI progress.

“自然语言是对世界的描述，代码是对 Solution 的描述” — the guest’s reason for treating code as a general digital-work substrate.

“模型可能就是新一代操作系统” — the platform thesis behind Model As Operating System.

“AI 取代的是不拥抱 AI 的人” — the episode’s practical advice about labor-market adaptation.

Connections

Contradictions