Researcher: GPT-6 Astra turned a month-long research job into one week, feels RSI momentum
i_dg23 · x · 2026-09-04
Researcher LiuZuxin shares hands-on experience handing real research jobs to GPT-6 Astra, arguing benchmarks are only a tiny part of the story. The post quotes OpenAI's official announcement that Astra achieves SOTA on Agents' Last Exam, AutomationBench, and ScreenSpot Pro.
- A research integration cycle that used to take at least a month of full-time work took just over a week, with only part-time steering
- Previous models required heavy prompting and skill-building; Astra largely just works
- Briefly dropping back to internal 5.6 made the intelligence gap obvious: you notice it when it's taken away
- The author directly feels strong RSI momentum — the task itself was improving the next model
More from AGI Musings
- EU will be AI's biggest victim, argues thread: it suppresses everything AI rewards — robleclerc · 2026-09-04
- Agents will soon build software faster than humans can review it — neurosp1ke · 2026-09-04
- François Chollet pulls AGI timeline to sooner than 2030 after Astra and ARC-AGI-3 — relegi · 2026-09-04
- HBR's 12,637-case study: a quarter of top AI uses now involve outsourcing your thinking — luisdans · 2026-09-04
- Study: AI companions rival human friendship, users mourn forced separations — EricTopol · 2026-09-04
- "Start preparing for the possibility that the world isn't going to end": a pushback on AI doom — BLUECOW009 · 2026-09-04