Key takeaways
• OpenAI claims GPT Astra 6 means “AGI is here,” but internal RSI is still at research-intern level (over half of 4–8h tasks need human intervention)—not a self-improvement breakthrough
• Real gains: ScreenSpot-Pro 76.9→92.7%, AutomationBench 18.1→41.4%, CAD +12.6pt; DeepSWE only +1.4pt—what’s sold is broader software/agent skill, not RSI-style AGI
• Industry trend: Fable and Gemini likewise stress automation/domain agents; after Coding Agent markets saturate, non-developer knowledge work explodes and Computer Use returns for visual/interactive verification
• Work traces are a training ore vein, but outside coding, env reconstruction, rewards, and credit assignment stay hard; experts + data firms dominate; Terminal-Universe / SDPO are only semi-automatic exits
• Whether “eating the world bite by bite” yields cross-domain general skills remains doubtful (generalization tax); the author argues commercial drive may outweigh AGI idealism
