The Superalignment team, established in July, is focused on the challenge of ensuring that AI remains secure and useful as it moves closer to and even surpasses human intelligence. "AGI is quickly getting closer," Leopold Aschenbrenner, the head of research at OpenAI and OpenAI, told WIRED. "We're likely to see superhuman-like models with vast capabilities and could be extremely dangerous, and we're not yet equipped with the methods to control these models."

The new research paper from OpenAI outlines an approach called supervision where a less advanced AI model guides the behavior of a higher-level one. This method is designed to safeguard the higher-end model's capabilities while ensuring it conforms to ethical and safe guidelines. This is considered to be a crucial step toward managing future superhuman AIs.

The tests involved using OpenAI's GPT-2 text generator to train GPT-4, a more advanced system. The researchers evaluated two ways to prevent the degradation in the performance of GPT-4. The first approach involved the training of progressively larger models, and the second one added an algorithmic modification to GPT-4. The latter was found to be more effective although the researchers acknowledge that the ability to control behavior perfectly isn't yet a certainty.

Industry response and future directions

Dan Hendryks, director of the Center for AI Safety, acknowledged the proactive approach of OpenAI in regulating superhuman AIs. In the Superalignment group's efforts are viewed as a significant initial step, but more research and development is needed to ensure the effectiveness of control systems.

OpenAI plans to dedicate the majority of its computing power to the Superalignment project and has launched a call for collaboration from other organizations. The company, in partnership together with Eric Schmidt, is offering 10 million dollars in grants to researchers working on AI control techniques. Furthermore, there will be a conference on superalignment next year to further explore this critical area.

Ilya Sutskever, a co-founder of OpenAI and a key player in the company's technological advances, co-leads the Superalignment team. Sutskever's involvement in the project is vital, particularly in light of the recent governance issues at OpenAI. Sutskever's knowledge and leadership skills are crucial in pushing this project to its next level.

The creation of methods to control super-intelligent AI is a difficult and urgent task. As AI technology rapidly advances, ensuring its alignment with human values and the safety of our children is becoming more important. The OpenAI initiative in this field is an important step but the journey towards reliable and effective AI-controlling systems remains still ongoing and requires the collaboration of the global AI research community.