Unlocking AI’s Creative Potential: The Emergence of AI Scientists Inventing and Experimenting at UBC

To review this article again, go to My Profile and then click on View saved stories.

An 'AI Researcher' Creates and Conducts Its Own Studies

Upon initial observation, a series of scholarly articles emanating from a leading artificial intelligence laboratory at the University of British Columbia in Vancouver may appear somewhat unremarkable. With their focus on slight advancements in current algorithms and concepts, they resemble the typical submissions found at an average AI symposium or publication.

However, the study stands out for its significance. This is attributed to the fact that it was solely conducted by an "AI scientist" created by the UBC laboratory in collaboration with experts from the University of Oxford and a startup named Sakana AI.

The initiative showcases a preliminary phase towards what could potentially be a groundbreaking maneuver: enabling AI to advance by generating and investigating new concepts. Currently, these concepts aren't exceptionally groundbreaking. Multiple studies detail modifications aimed at enhancing a technique for creating images called diffusion modeling; another study presents a strategy for accelerating the learning process in deep neural networks.

"Jeff Clune, the head of the UBC lab, acknowledges that these concepts aren't revolutionary or exceptionally innovative. However, he believes they're interesting ideas worth exploring."

Despite the remarkable capabilities of current AI technologies, their reliance on data created by humans restricts their potential. Should AI systems evolve to discover and investigate novel concepts through autonomous exploration, they could surpass the limits imposed by human-derived knowledge.

Clune's research group has been at the forefront of developing AI systems capable of learning through innovative methods. One notable venture, known as Omni, focused on creating complex behaviors for virtual entities across various simulated environments. This initiative involved identifying compelling behaviors and refining them through continuous iteration. Traditionally, these AI systems relied on manually programmed criteria to determine what constituted as interesting behavior. However, the advent of advanced language models has revolutionized this process by equipping AI with the capacity to discern fascination similarly to human thought processes. In a more recent endeavor, Clune's team leveraged this methodology to empower AI to autonomously generate code. This code enables virtual characters to perform a diverse array of activities within a digital realm reminiscent of Roblox.

In Clune's laboratory, the AI scientist serves as a demonstration of exploring potential innovations. This software generates ideas for machine learning projects, selects the most viable ones with assistance from a large language model (LLM), and then proceeds to code and execute these ideas in a continuous cycle. Although the outcomes have so far been modest, Clune believes that programs focused on open-ended learning, similar to language models, could significantly improve in performance with increased computational resources.

"Clune describes the potential of LLMs as akin to venturing into an uncharted land or an undiscovered world," he states. "The future discoveries remain a mystery, yet every direction promises something novel."

Tom Hope, who serves as an assistant professor at the Hebrew University of Jerusalem and also works as a research scientist at the Allen Institute for AI (AI2), believes that AI scientists, including large language models (LLMs), tend to be overly imitative and shouldn't be viewed as dependable. "At this moment, none of the elements can be deemed reliable," he states.

Hope highlights that the initiative to incorporate automation into scientific research dates back to the pioneering efforts of AI experts Allen Newell and Herbert Simon in the 1970s, followed by the contributions of Pat in subsequent years.

At the Institute for the Study of Learning and Expertise, Langley highlights that multiple research collectives, such as a group from AI2, have also utilized large language models (LLMs) recently for tasks like formulating hypotheses, composing academic papers, and evaluating studies. "They've really tapped into the current spirit," comments Hope on the efforts of the UBC team. "Naturally, the trajectory they're on holds immense value, potentially."

The possibility of LLM-based systems generating genuinely innovative or groundbreaking concepts is still uncertain. "That's the question worth a trillion dollars," notes Clune.

Without the need for groundbreaking discoveries, continuous education could be crucial for creating AI systems that are both more skilled and practical in the present moment. A recent publication by Air Street Capital, a venture capital company, underscores the significance of Clune's research in advancing AI agents—software that can independently carry out beneficial operations on computers. It appears that major AI corporations are unanimously considering these agents as the next major breakthrough.

This week, Clune's research team unveiled their newest initiative in exploratory learning: a software that creates and develops AI entities. These entities, designed by the AI, show superior performance over those created by humans in certain areas, including mathematics and understanding written material. The upcoming challenge for the team is to develop strategies to ensure these systems do not create malfunctioning agents. "This area of work carries potential risks," Clune comments. "It's crucial we approach it correctly, though I believe it's achievable."

Explore Further …

Dive into the World of Politics: Subscribe to our newsletter and tune into our podcast

Exploring the effects of distributing unconditional cash to individuals

Not everyone experiences weight loss with Ozempic.

The Pentagon is planning to allocate $141 billion towards a machine designed for apocalyptic scenarios.

Attend the Energy Tech Summit happening on October 10th in Berlin.

Additional Content from WIRED

Critiques and Manuals

© 2024 Condé Nast. All rights are reserved. WIRED may receive a share of revenue from products bought via our website, as a part of our Affiliate Partnership agreements with retail sellers. Content from this site must not be copied, shared, transmitted, stored, or used in any form without explicit written consent from Condé Nast. Choices regarding advertisements.

Choose a global website