OpenAI announces new team to control "super-intelligent" artificial intelligence

News on July 6, on Wednesday, local time in the United States, artificial intelligence startup OpenAI announced that it is forming a new team to develop guidance and control of "superintelligent" ) approach to artificial intelligence systems. The team is led by Ilya Sutskever, chief scientist and co-founder of OpenAI.

In a blog post, Suskwell and Jan Leike, head of the OpenAI coordination team, predicted that artificial intelligence with greater intelligence than humans will emerge within a decade. They warn that such artificial intelligence will not necessarily be benevolent towards humans, so it will be necessary to research ways to control and limit it.

Suskwell and Lake wrote: "Currently, we do not have any solutions that can be used to manipulate or control a potentially superintelligent AI and prevent it from getting out of hand. Our current techniques for tuning AI, such as Reinforcement learning from human feedback relies on the ability of humans to supervise artificial intelligence. But humans will not be able to reliably supervise artificial intelligence systems that are much smarter than us." ), OpenAI is creating a new super-aligned team, co-led by Susqueville and Lake, which will have access to up to 20% of OpenAI’s computing resources. This team will work over the next four years to solve the core technical challenges of controlling superintelligent artificial intelligence, along with scientists and engineers from OpenAI's former Alignment Division, as well as researchers from other parts of the company.

So, how do we ensure that artificial intelligence systems that are much smarter than humans follow human intentions? This will require the help of what Suskville and Lake describe as "human-level automated alignment researchers." The high-level goals are to use human feedback to train AI systems, train AI to assist in the evaluation of other AI systems, and ultimately build AI that can conduct alignment studies. “Alignment research” here refers to ensuring that the artificial intelligence system achieves the expected results or does not deviate from the research track.

OpenAI’s hypothesis is that artificial intelligence can perform alignment research faster and better than humans.

"As we make progress in this area, our artificial intelligence systems can take over more and more of the alignment work and ultimately conceive, implement, research and develop better alignment technologies than we have today," Lake and colleagues John Schulman and Jeffrey Wu hypothesized in a previous blog post. "They will work with humans to ensure that their own successors are always aligned with humans. Human researchers will increasingly focus on reviewing alignment studies done by AI systems, rather than conducting such studies themselves Research."

Of course, no method is foolproof, and Lake, Schulman, and Jeffrey Wu acknowledge in their article that OpenAI's approach has many limitations. They say using AI for assessment has the potential to amplify inconsistencies, biases or vulnerabilities in AI. It may turn out that the most difficult part of the alignment problem may have nothing to do with engineering at all.

But Suskville and Lake decided it was worth a try. They write: "Superintelligent alignment is fundamentally a machine learning problem, and we believe that good machine learning experts (even if they haven't started working on alignment) will be critical to solving this problem. We plan to share this widely The results of our efforts and regard promoting the consistency and security of non-OpenAI models as an important part of our work."

The above is the detailed content of OpenAI announces new team to control "super-intelligent" artificial intelligence. For more information, please follow other related articles on the PHP Chinese website!