At first glance, a recent batch of research papers produced by the University of British Columbia’s preeminent artificial intelligence lab in Vancouver may not seem all that significant. Containing incremental improvements to existing algorithms and ideas, they read like the content of your average AI conference or journal.
But the research is actually extraordinary. That’s because it’s entirely the work of “Artificial Intelligence Scientist“developed in a UBC lab together with scientists from the University of Oxford and a startup called Conversational AI.
This design shows an early step toward what could be a revolutionary trick: enabling AI to learn by inventing and exploring novel ideas. They just aren’t supernovae at the moment. Several papers describe modifications to an image-generating technique known as diffusion modeling; another outlines an approach to accelerating learning in deep neural networks.
“These aren’t groundbreaking ideas. They’re not wildly creative,” he admits. Jeff Cluneprofessor who directs the UBC lab. “But they seem like pretty cool ideas that someone could try.”
As amazing as today’s AI programs can be, they are narrow by the need to consume training data generated by humans. If AI programs can learn in an open way, by experimenting and exploring “interesting” ideas, they can unlock possibilities beyond anything humans have shown them.
Clune’s lab has previously developed AI programs designed to learn this way. For example, one program called Omni attempted to generate behavior for virtual characters in several video game-like environments, setting aside those that seemed intriguing and then repeating them with novel designs. Previously, these programs required hand-coded instructions to define attractiveness. Huge language models, however, provide a way to enable these programs to identify what is most intriguing. Another last project from Clune’s lab used this approach to enable AI programs to invent code that would allow virtual characters to perform various actions in a Roblox-like world.
The AI scientist is one example of Clune’s lab experimenting with possibilities. The program comes up with machine-learning experiments, decides what seems most promising with the aid of LLM, and then writes and runs the necessary code—rinse and repeat. Despite the disappointing results, Clune says that open-source learning programs, like the language models themselves, could become much more productive as the computing power that powers them increases.
“It’s like discovering a new continent or a new planet,” says Clune of the possibilities that an LLM opens up. “We don’t know what we’ll discover, but everywhere we turn, there’s something new.”
Volume Hopeassistant professor at the Hebrew University of Jerusalem and research scientist at the Allen Institute for AI (AI2), says that the AI Scientist, like the LLM, appears to be highly derivative and cannot be considered trustworthy. “None of the components are trustworthy at this time,” he says.
Hope notes that efforts to automate elements of scientific discovery have been going on for decades, led by AI pioneers Allen Newell AND Herbert Simon in the 1970s and later work Pat
Langley at the Institute for the Study of Learning and Expertise. He also notes that several other research groups, including a team at AI2, have recently used the LLM to aid generate hypotheses, write papers and peer-review research. “They’ve captured the zeitgeist,” Hope says of the UBC team. “The direction is obviously incredibly valuable, potentially.”
It also remains unclear whether LLM-based systems will ever be able to come up with truly novel or groundbreaking ideas. “That’s the trillion-dollar question,” Clune says.
Even without scientific breakthroughs, open learning could prove crucial to developing more productive and useful AI systems now. Report published this month by Air Street Capital, an investment firm, highlights the potential of Clune’s work to develop more powerful and reliable AI agents—programs that autonomously perform useful tasks on computers. All the gigantic AI companies seem to see agents as the next gigantic thing.
This week, Clune’s lab revealed its latest open learning project: An AI program that invents and builds AI agents. AI-designed agents outperform human-designed agents in some tasks, such as math and reading comprehension. The next step will be to figure out how to prevent such a system from generating agents that behave badly. “That’s potentially dangerous,” Clune says of the work. “We have to get it right, but I think it can be done.”