Following the controversy surrounding Sam Altman’s departure and return to OpenAI, the firm’s Superalignment team remains dedicated to tackling the challenge of controlling artificial intelligence (AI) beyond human capabilities. Led by co-founder and chief scientist Ilya Sutskever, the team is actively developing strategies for governing and regulating superintelligent AI systems.
TechCrunch has reported that the Superalignment team recently presented its latest work at the annual machine learning conference NeurIPS in New Orleans. The team uses an analogy where a less advanced AI model, such as GPT-2, guides a more advanced model, like GPT-4, towards desired outcomes while avoiding undesirable ones. This approach is crucial in achieving alignment goals for superintelligent AI.
In a recent study, OpenAI trained its GPT-2 model to perform various tasks, including chess puzzles and sentiment analysis. The responses generated by GPT-2 were then used to train the GPT-4 model. According to Tech.co, GPT-4 demonstrated a 20-70% performance improvement compared to GPT-2, showcasing its superior abilities. However, the results fell short of GPT-4’s full potential.
The experiment also highlighted the concept of ‘weak-to-strong generalization,’ where GPT-4 demonstrated the ability to avoid many of the mistakes made by GPT-2. This phenomenon suggests that future AI models might be better at spotting dangerous actions that could cause significant harm.
Despite favorable results, GPT-4’s performance was still hindered after training with GPT-2. This emphasizes the need for further research before considering humans as suitable supervisors for more advanced AI models.
OpenAI has consistently expressed concerns about the potential dangers of superintelligent AI. Co-founder Ilya Sutskever stressed the importance of preventing AI from turning rogue, acknowledging the risks it poses, including the disempowerment or even extinction of humanity if left unchecked.
To promote research in this area, OpenAI has introduced a $10 million grant program for technical research on superintelligent alignment. The program is open to academic labs, nonprofits, individual researchers, and graduate students.
Eric Schmidt, former Google CEO and chairman, renowned for his advocacy of AI doomism, is among the notable funders of the grant program. Schmidt emphasizes the importance of aligning AI with human values and is proud to support OpenAI’s responsible development and control of AI for the benefit of society.
In light of the rapid development of AI, Pope Francis recently cautioned global leaders about the risks associated with uncontrolled technological progress. While acknowledging the benefits of scientific advancements, the Pope expressed concern about the unprecedented control these advances may exert over reality, potentially jeopardizing humanity’s survival.
The pontiff urged leaders to carefully examine the intentions and interests of AI developers, cautioning against selfish motives. He emphasized the need to direct research towards peace, the common good, and integral human development, warning against the potential exploitation of AI efficiencies that could lead to unclear decision-making criteria and hidden obligations.
As the world grapples with the challenge of controlling superintelligent AI, OpenAI’s Superalignment team continues to make significant progress in addressing this complex issue. With ongoing research, collaborations, and grant programs, the aim is to ensure that AI development remains aligned with human values, promoting the responsible and beneficial use of this powerful technology.
