OpenAI is also seeking external assistance, offering $10 million in grants in partnership with Eric Schmidt, former CEO of Google, to researchers who can advance topics such as weak-to-strong supervision and interpretability of advanced models. The company plans to hold a conference on superalignment next year. The question of controlling future AI technology has gained prominence due to the rapid advancement of AI, exemplified by OpenAI's ChatGPT.
Key takeaways:
- OpenAI has established a Superalignment research team to develop methods to control superhuman AI models, dedicating a fifth of its computing power to the project.
- A research paper released by OpenAI details experiments on allowing a less intelligent AI model to guide a smarter one, a process they believe will be crucial as AI becomes more powerful than humans.
- The Superalignment team is co-led by Ilya Sutskever, an OpenAI co-founder and chief scientist, who has been a significant driving force in the project.
- OpenAI is offering $10 million in grants, in partnership with Eric Schmidt, to outside researchers to advance topics including weak-to-strong supervision, interpretability of advanced models, and strengthening models against prompts designed to break their restrictions.