In Short:
Researchers at the University of British Columbia, collaborating with Oxford and Sakana AI, developed an AI program that can autonomously create and test machine learning experiments. While the recent research shows modest improvements in AI techniques, it signifies a step toward AI that learns through exploration rather than just human data. This could lead to more powerful AI systems, though concerns about reliability and safety remain.
A recent collection of research papers from a leading artificial intelligence lab at the University of British Columbia in Vancouver presents findings that, at first glance, may appear unremarkable. With enhancements to already established algorithms, the papers resemble the type of work typically showcased at an average AI conference or journal.
Noteworthy Contributions
However, this research holds significant implications. It stems entirely from an “AI scientist” developed at the UBC lab in partnership with researchers from University of Oxford and the startup Sakana AI.
Exploration of AI Learning
The project marks an initial advancement toward a potentially transformative approach: allowing AI to autonomously learn through the invention and exploration of novel ideas, even if these ideas are not groundbreaking at present. The papers include modifications to enhance a technique known as diffusion modeling for image generation and propose strategies to accelerate learning in deep neural networks.
Perspectives from Experts
“These are not breakthrough ideas. They’re not wildly creative,” acknowledges Jeff Clune, the professor leading the UBC lab. “But they seem like pretty cool ideas that somebody might try.”
Current AI models exhibit remarkable capabilities but are constrained by their reliance on human-generated training data. If AI systems can learn through open-ended experimentation driven by “interesting” concepts, they may achieve abilities beyond those demonstrated by humans.
Previous Developments in AI Learning
Clune’s lab has already produced programs designed for this type of learning, such as Omni, which generates behaviors for virtual characters within various video game-like settings, retaining those it finds engaging and iterating on them with new designs. Unlike earlier iterations that required manual programming to define “interestingness,” large language models (LLMs) now enable these programs to autonomously determine what is intriguing. A recent project illustrated this capability by allowing AI systems to create code for virtual characters in a Roblox-like environment.
Future Prospects and Challenges
The AI scientist program exemplifies how Clune’s lab is exploring future possibilities. This program generates machine learning experiments and, with the assistance of an LLM, determines the most promising avenues, subsequently creating and executing the necessary code—a continuous cycle of innovation. Despite the modest outcomes thus far, Clune asserts that open-ended learning programs, much like the LLMs themselves, could significantly enhance their performance as computational power scales up.
“It feels like exploring a new continent or a new planet,” Clune reflects on the potential discoveries enabled by LLMs. “We don’t know what we’re going to discover, but everywhere we turn, there’s something new.”
Concerns and Historical Context
Tom Hope, an assistant professor at the Hebrew University of Jerusalem and a research scientist at the Allen Institute for AI (AI2), expresses caution regarding the AI scientist, noting its highly derivative nature and questioning its reliability: “None of the components are trustworthy right now,” he remarks.
Hope highlights that attempts to automate aspects of scientific discovery have roots in efforts dating back to the 1970s by AI pioneers like Allen Newell and Herbert Simon, as well as more recent endeavors led by Pat Langley at the Institute for the Study of Learning and Expertise. He also points out that several teams, including AI2, have utilized LLMs for generating hypotheses and reviewing research. “They captured the zeitgeist,” Hope assesses the UBC team’s efforts, adding that this direction is undoubtedly valuable and potentially groundbreaking.
Ongoing Questions and Investment Potential
The prospect of LLM-based systems generating genuinely novel or transformative ideas remains uncertain. “That’s the trillion-dollar question,” Clune remarks on this challenge.
Despite the absence of immediate scientific breakthroughs, open-ended learning could be crucial for developing more capable and effective AI systems in the near future. A report released this month by Air Street Capital, an investment firm, underscores the potential of Clune’s work in creating more advanced and reliable AI agents—programs designed to autonomously execute beneficial tasks on computers. Major AI companies are increasingly viewing agents as the next frontier in AI development.
Latest Advancements
This week, Clune’s lab unveiled its latest initiative in open-ended learning: an AI program designed to invent and build AI agents. Remarkably, these AI-generated agents have shown superior performance to their human-designed counterparts in areas like mathematics and reading comprehension. Moving forward, the challenge lies in ensuring that this system does not produce agents that engage in harmful behavior. “It’s potentially dangerous,” Clune cautions regarding this line of research. “We need to get it right, but I think it’s possible.”
