OpenAI, the creator of ChatGPT, has announced plans to dedicate a fifth of its compute power to prevent the potential dangers posed by rogue artificial intelligence (AI). The start-up, which is backed by Microsoft, aims to solve the alignment problem by ensuring that the goals of AI systems are beneficial to humans. In a recent blog post, Ilya Sutskever and Jan Leike, co-founders of OpenAI, expressed concerns about the immense power of superintelligence and its potential to disempower humanity or even lead to human extinction. They emphasized the need for breakthroughs in alignment research to control superintelligent AI.
Sutskever and Leike predicted that superintelligent AI systems, which possess greater intelligence than humans, could emerge within this decade. They stressed the importance of developing superior techniques to control and steer such AI systems, hence the focus on alignment research. OpenAI plans to dedicate 20% of its compute power over the next four years to tackle this challenge. Additionally, the company will establish a new research team, the Superalignment team, to lead this effort.
The primary objective of the team is to create an AI alignment researcher capable of human-level understanding, with the subsequent goal of scaling up this capability using vast amounts of compute power. OpenAI intends to train AI systems using human feedback, enable them to assist in human evaluation, and eventually have AI systems conduct alignment research themselves.
However, some AI safety advocates, such as Connor Leahy, caution that OpenAI’s approach is flawed. Leahy believes that solving the alignment problem should be a priority before pursuing human-level AI, as the initial AI systems may behave unpredictably and cause havoc if not properly controlled. He argues against relying on a plan that could potentially lead to unintended and unsafe consequences.
Concerns about the risks associated with AI have been a prominent topic among AI researchers and the general public. In April, a group of industry leaders and experts in AI signed an open letter calling for a six-month pause in the development of AI systems more powerful than OpenAI’s GPT-4, citing potential societal risks. Additionally, a Reuters/Ipsos poll in May revealed that over two-thirds of Americans are worried about the potential negative effects of AI, with 61% believing it could pose a threat to civilization.
OpenAI’s commitment to dedicating compute power and resources to the alignment problem signifies the importance of addressing the potential dangers of superintelligent AI. Their efforts aim to ensure that AI remains beneficial and controllable for humanity, but critics argue for a more cautious and proactive approach to mitigate unintended consequences. The future of AI hinges on finding solutions that prioritize the alignment of AI goals with human values.
