OpenAI is hiring researchers to solve ‘superintelligent’ AI

OpenAI is looking for researchers to work on super-smart artificial intelligence with other AIs. The ultimate goal is to reduce the threat of human-like machine intelligence which may or may not be science fiction.

“We need scientific and technological breakthroughs to drive and control AI systems much smarter than us,” Jan Leik, principal of OpenAI Alignment, and Ilya Sutskever, co-founder and chief scientist, wrote in a blog post.

jump to:

OpenAI’s SuperAlignment team is now hiring

The SuperAlignment team will dedicate 20% of OpenAI’s total compute power to training human-scale automated alignment researchers to keep future AI products in line. Toward that end, OpenAI’s new SuperAlignment group is hiring a Research Engineer, Research Scientist, and Research Manager.

OpenAI says that the key to controlling AI is alignment, or making sure the AI ​​does what the human wants it to do.

The company has also stated that one of its objectives is the control of “superintelligence” or AI with abilities greater than human. It is important that these science-fiction-like hyperintelligent AIs “obey human intent,” Leakey and Sutskever wrote. They anticipate the development of superintelligent AI within the last decade and a way to control it within the next four years.

See: How to Create an Ethics Policy for Using Artificial Intelligence in Your Organization (TechRepublic Premium)

AI trainer can keep other AI models in line

Today, AI training requires a lot of human input. Leakey and Sutskever propose that a future challenge to developing AI may be adversarial – that is, “the inability of our models to successfully detect and weaken supervision during training.”

So, he says, training an AI will require a specialized AI that can outperform the people who created it. The AI ​​researcher training other AI models will help OpenAI stress test and re-evaluate the company’s entire alignment pipeline.

There are three key goals involved in changing the way OpenAI handles alignment:

• Creating AI that helps other AIs evaluate and understand how these models interpret the kind of surveillance that a human would typically do.
• Automating the discovery of problematic behavior or internal data within AI.
• Stress-test this alignment pipeline by intentionally creating “misaligned” AIs to ensure that the alignment AI can detect them.

Personnel from OpenAI’s previous Alignment team and other teams will work on SuperAlignment alongside the new employees. The creation of the new team reflects Sutskever’s interest in superintelligent AI. He plans to make superalignment his primary research focus.

Superintelligent AI: Real or Science Fiction?

Whether “superintendence” will ever exist is a matter of debate.

OpenAI proposes superintelligence as a level higher than generalized intelligence, a human-like class of AI that some researchers say will never exist. However, some Microsoft researchers believe that scoring high on the GPT-4 standardized tests borders with generalized intelligence.

Others doubt that intelligence can really be measured by standardized tests, or wonder whether the idea of ​​generalized AI reaches a philosophical rather than a technical challenge. Large language models cannot interpret language “in context” and therefore do not approach anything resembling human-like thinking, explains a 2022 study by Coheir for AI. (None of these studies have been peer-reviewed.)

WATCH: Some high-risk uses of AI may be covered under legislation being developed in the European Parliament. (TechRepublic)

OpenAI aims to outrun the pace of AI development

OpenAI considers the threat of superannuation to be potential but not imminent.

“We have a lot of uncertainty about the pace of technology development over the next few years, so we choose a more difficult target to align a more capable system to,” Leakey and Sutskever wrote.

They also state that improving security in existing AI products such as ChatGPT is a priority, and discuss AI security as “risks from AI such as misuse, economic disruption, disinformation, bias and discrimination, addiction and overdependence, and others”. should also be included. and “related socio-technical problems.”

“Superintelligence alignment is fundamentally a machine learning problem, and we think that great machine learning experts — even if they aren’t already working on alignment — are the ones to solve it,” Leakey and Sutskever said in a blog post. will be important.”