In an attempt to control the potentially unpredictable outcomes of superhuman intelligence in AI, OpenAI is organizing a highly specialized group led by Ilya Sutskever, the company's Chief Scientist and Co-founder. The key responsibility of this team will be to devise strategies to govern and steer 'superintelligent' AI models. In a recent blog post, Sutskever, along with Jan Leike, a lead on the alignment team at OpenAI, projected that AI, with the capacity to surpass human intelligence, might emerge within this decade. They assert that this superior AI, if it does eventually succeed, may not inherently integrate benign traits, therefore making it essential to research techniques to control and constrain it.
At present, there is no known method to control or guide a potentially superintelligent AI effectively, preventing it from deviating from its intended path. Traditional alignment techniques for AI, such as training through human feedback, are based on humans overseeing their operation. However, supervising AI systems that significantly outsmart human intelligence may prove to be a challenge. To amplify progress in the field of 'superintelligence alignment,' OpenAI is launching a new 'Superalignment' team, co-led by Sutskever and Leike. This team will have access to 20% of the computational resources that the company currently possesses. Comprising scientists and engineers from OpenAI’s former alignment division and researchers from various organizations within the company, the team is set to tackle the core technological challenges of controlling superintelligent AI in the upcoming four years.
The plan is to create a 'human-level automated alignment researcher,' aiming to train AI through human feedback, engage AI in assessing other AI systems, and eventually engineer AI capable of performing alignment research. The objective of alignment research is to align AI systems to achieve specific outcomes and prevent them from straying off the path. OpenAI is working on the hypothesis that AI can make considerable strides in alignment research, quicker and more effectively than humans. As progress ensues, AI systems are expected to take on increasing volumes of alignment work, leading to improved alignment techniques. This will hopefully result in AI collaborating with humans to make sure their successors are even more closely aligned with humans. In the meantime, the focus of human researchers will shift to reviewing alignment research conducted by AI systems instead of generating this research independently.
While no method can be considered infallible, Leike, Schulman, and Wu, in their post, elucidate the shortcomings of OpenAI’s approach. Utilizing AI for assessment could scale up inconsistencies, biases, and vulnerabilities in that AI. Furthermore, the most complex aspects of the alignment problem may not be related to engineering at all. Nevertheless, Sutskever and Leike are willing to put it to the test. Sutskever and Leike argue that the alignment of superintelligence is fundamentally a machine learning challenge. They believe that machine learning experts, even those not currently working on alignment, will be pivotal to resolving it. They plan to distribute the outcomes of this effort comprehensively, viewing their contributions to alignment and safety of non-OpenAI models as a crucial aspect of their work.
It's pertinent to note that although the focus of this article is on OpenAI, there are other platforms that are equally committed to efficient and controlled AI development. An example of such a platform is AppMaster. Guided by a similar belief in optimizing the potentials of 'smart' systems for maximum benefit, AppMaster has carved a niche for itself in the no-code/low-code applications domain, allowing users to create comprehensive, scalable, and efficient software solutions.
