AGI Three Acts

AGI Three Acts is the route map in 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: first ChatGPT-style chatbots, then coding agents such as Claude Code and Codex, then automated AI researchers. The episode treats coding not as a narrow developer market, but as the second act of AGI because code can express and execute solutions across much of digital knowledge work.

The third act connects to ML Coding, AI For Science, Discovery Model, and Recursive Self-Improvement without assuming fully autonomous self-improvement has already arrived. The source’s more conservative version is that coding agents shorten research, data, and experiment loops enough that humans can test more ideas and eventually delegate more AI-research work.

171: 【AI季报 26Q2】从 coding 到 RSI，强者愈强的未来？ makes the third act more explicit through Auto Research and RSI. It treats coding as both a current revenue/product wedge and the substrate for automated research: agents that write and run experiments can become part of a loop that improves future model training, benchmarks, and tools.

Key Claims

• Chatbots made AI broadly accessible, but left most valuable work inside conversation.
• Coding agents turn model capability into executable digital work because code has tests, logs, errors, repositories, and reviewable artifacts.
• The second act can accelerate the third by making AI research itself faster: experiment code, data processing, evaluation, and multimodal iteration can be compressed from weeks to days.
• The route depends on AI Coding Verification, Research Taste, and Problem Definition In Research because faster generated experiments can still be wrong, irrelevant, or poorly measured.
• Agent Harness is part of the second act because agents need environments, tools, permissions, memory, and feedback to do work rather than only answer.
• The social consequence is Intelligence Devaluation: if large portions of knowledge work become executable through agents, junior white-collar and outsourcing work may be repriced.
• The LateTalk Q2 source adds that coding is also a strategic self-improvement input because model teams can use coding agents to accelerate their own research and engineering loops.

Connections

• ChatGPT — first-act chatbot and consumer-assistant symbol.
• Claude Code, Codex, Gemini CLI, and Vibe Coding — second-act coding-agent and AI-coding surfaces.
• ML Coding, AI For Science, Recursive Self-Improvement, and Discovery Model — third-act research automation branch.
• Agent Harness, Model Harness Co-Evolution, and Agent Post-Training — infrastructure and training loops that make agent work improve models.
• AI Programming Engine Shift, AI Coding Verification, and AI Engineering Thinking — coding-work mechanisms and verification constraints.
• Model As Operating System, AI Investment Metrics, and AI Commercialization Pressure — strategic and investment implications of the route map.
• Auto Research, GPT-5.6, Fable 5, and Recursive — Q2 2026 update from LateTalk on the transition from coding to RSI.