OpenAI is recruiting for a new research team focused on multi-agent systems, signaling the company's push towards the next phase of AI development.
Noam Brown, an OpenAI researcher, announced on X that the company is seeking to expand a new multi-agent research team. The development of AI agents aligns with the third level of OpenAI's recently introduced five-level scale for measuring progress towards AGI. OpenAI currently considers itself at the threshold of the second stage, known as "reasoners," with its o1 model.
According to The Information, OpenAI has been working on two types of AI agents for automating complex tasks. One type aims to control devices to transfer data between documents or complete expense reports. The other focuses on web-based tasks such as gathering public data or booking flights.
Google DeepMind is also prioritizing AI agent development. CEO Demis Hassabis expects such systems could be ready for use within one to two years. In a Bloomberg interview, he said: "I'm really excited about the next stage of these large general models. You know, I think the next things we're going to see perhaps this year, maybe next year, is more agent like behavior."
Nvidia CEO Jensen Huang shares this view. In a conversation with Salesforce CEO Marc Benioff, he predicted that AI agent development will be "spectacular and surprising" in the next one to two years.
Multi-agent reasoning
While projects using LLM agents already exist, such as in translation, the planned work on multi-agent systems aims higher. OpenAI and Brown see multi-agent systems as a path to improved AI capabilities. Brown stated on X: "We view multi-agent as a path to even better AI reasoning."
Brown, formerly at Facebook AI Research, is known for his work on AI systems that beat human players in complex games like poker and Diplomacy. His research on systems like the poker AI Libratus explored "test-time compute" - using more computing time for answers led to better moves. He joined OpenAI to apply his work to language models. OpenAI's latest o1 model is a result of this work.
Interested candidates can apply through OpenAI's online form.
