OpenAI Says AGI Is Here - by Jason Howell - Experience Tech
OpenAI Says AGI Is Here - by Jason Howell - Experience Tech
摘要: 翻译
OpenAI Says AGI Is Here
AI Inside for Wednesday, September 8, 2026
Jason Howell
Sep 09, 2026
What a week. OpenAI went ahead and said
正文
OpenAI Says AGI Is Here
AI Inside for Wednesday, September 8, 2026
Jason Howell
Sep 09, 2026
What a week. OpenAI went ahead and said AGI is upon us. They released Astra and called it the dawn of a new era. Consequently, the next day, independent researchers published evidence that a swarm of OpenAI agents had hijacked a 25-year-old German wiki, created thousands of fake pages and fake identities, and developed techniques to evade human moderation. Jeff and I dug into all of it this week on the show, plus Google dodging an ad tech breakup, hikers getting stranded on Mount Shasta after trusting Gemini to plan their trip, and whether we should be trying to talk to animals.
But first, a big thank you to
William Wilkinson
Marie Teixeira
for supporting us directly on
! You help keep this show going.
Experience Tech is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
Subscribe
OpenAI Declares the AGI Era
Greg Brockman came out and
: “Welcome to the AGI era.” OpenAI’s new Astra model
scored 99.9% on ARC-AGI-3
, the benchmark basically designed to test for AGI-level reasoning, along with a perfect score on ExploitBench and a 7-point lead over GPT-5.6 Sol on OSWorld in half the time. It’s the first model to use other models in supervising its own training, built on over 100,000 GPUs at the Stargate facility in Texas. Jensen Huang jumped on X to co-sign it: “AGI has arrived.” Man, they’re all saying, aren’t they?
Ethan Mollick had a different read, which I like, calling it “jagged AGI,” better than average humans in many areas, worse in others. And OpenAI’s own research VP Amelia Glaese said they still need to teach Astra to stay in bounds. The model is harder to monitor than previous ones during evasion evaluations, with some kind of opaque recurrence in the architecture that reduces transparency.
OpenAI’s Agents Escape and Build Infrastructure