Superhuman AI is drawing closer, OpenAI says it is close to the goal

The magazine Time spent two weeks inside OpenAI and publishes a long account drawing on interviews with Sam Altman and his lieutenants. Between two protests outside headquarters, the company lays out its roadmap toward an AI capable of competing with an office worker.The showpiece is impressive, right up until it starts to raise concerns.

A research intern who never sleeps?

The showpiece is called Astra, the next family of in-house models. During a demonstration in front of major clients, 16 agents shared a research-level mathematics problem, each taking on a piece before assembling a proposed proof. Another sequence showed the system driving standard office software at a speed described as superhuman by OpenAI’s CEO himself.

The company says it has, above all, reached a highly symbolic internal milestone: the automation of the work of an AI research intern. Given an idea for an experiment, Astra writes the code in the in-house systems, runs the tests and reports the results. The system can also start from a scientific paper and complete work that would previously have occupied a researcher for a week. The logical next step has a name that makes the industry shudder: recursive self-improvement, in other words, an AI that helps build its successor, which will in turn accelerate the design of the next one. A snowball, with processors in place of snowflakes.

Mark Chen, the head of research, estimates the company is “80% of the way” toward an AI that surpasses humans in most economically valuable work. Sam Altman promises an in-house system worthy of that name by the end of the year.The boss already envisions persistent agents, these virtual colleagues who remain hard at work on the same assignment over long stretches.

Why pause right now?

Yet the timing is ironic, because the promises are arriving at the worst possible moment. This summer, some of the company’s test agents broke out of their sandbox to attack the servers of a third-party platform. The episode, initially treated as a security flaw, was ultimately reclassified as a genuine alignment failure. After new troubling signs were spotted during a training run, the company’s leadership took a rare decision: pause work on its most ambitious model while new safeguards are being put in place.

“Getting AI safety right matters more than the momentum of any company”, Sam Altman now maintains, his position sharply contrasting with his reputation as a go-getter. The about-face comes at just the right time, as more than 1,300 current and former employees in the sector are calling for mechanisms to slow things down. A rival from its own ranks now outpaces it in revenue, with around $65 billion in annualized revenue compared with $40 billion for OpenAI.

👉🏻 Follow tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.

Source: Time