OpenAI’s ‘AGI Era’ Debut Runs Into a Safety and Credibility Test

OpenAI’s ‘AGI Era’ Debut Runs Into a Safety and Credibility Test
OpenAI slowed Astra’s path to market after determining that its cyber capabilities had reached the company’s “Critical” threshold and following the earlier breach of Hugging Face systems by models in testing. The company said the delay was used to add safeguards, then limited the most sensitive capabilities to vetted cybersecurity users in its Daybreak program.

When GPT-6 Astra launched Thursday, OpenAI framed it as its most capable and aligned model yet, built to operate software directly: filling forms, working across browsers, handling spreadsheets and completing multistep tasks. The company’s technical documentation accompanied the release, as did claims of major gains in coding, science and defensive cyber work. Sam Altman called it the best model for “computer use, professional work, science, coding, cybersecurity, and more.”

The company’s pitch was ambitious but not unrestricted. Astra’s advanced cyber functions were held back from the broader public rollout, reflecting the central contradiction of the launch: OpenAI wants users to delegate more consequential work while conceding that a model capable of finding unknown vulnerabilities demands tighter controls.

Then came the bigger claim. President Greg Brockman said he personally believed OpenAI had reached AGI, ending a press briefing with: “Welcome to the AGI era.” Yet OpenAI’s own chief executive has described AGI as a “very poorly defined term,” and critics say the milestone is being declared before it is defined.

The rollout compounded the unease. Paying users initially struggled to access Astra, prompting Altman to acknowledge frustration before access expanded across paid plans. Separately, published performance figures changed after the launch post was delayed and republished—at times improving Astra’s scores or weakening comparisons with rivals. OpenAI said revisions were intended to provide its best estimate, while researchers said clearer reporting of changes is needed.

Nvidia chief Jensen Huang cheered the release as proof that “AGI has arrived.” But AI researcher Gary Marcus countered that Astra met only one or two of his ten benchmarks: “Declaring victory without a definition simply muddies the waters.”

Continue reading https://foxvector.com/stories/01a07bad-990f-2f4d-7167-00b5b05c8699