AiWallz and Open-endedness
Posted on July 1, 2024
We wanted to build a tool that can automatically generate wallpapers most likely to be preferred by users. However, we don't want to limit them to the same old wallpapers they have seen before. Instead, we aim to provide a wide variety of interesting wallpapers that they might still like. By borrowing concepts from the AI research field of open-endedness, we can automatically generate wallpapers that are novel, interesting, and tailored to a user's preferences.
What is open-endedness in AI?
Open-ended AI algorithms aim to continuously invent new and ever-more complex tasks and solve them continually, even endlessly. More formally, a recent position paper on how Openedness is Essential for Artificial Superhuman Intelligence defines that "From the perspective of an observer, a system is open-ended if and only if the sequence of artifacts it produces is both novel and learnable." The concept of open-endedness often reflects the way human innovation and discovery work, with each breakthrough leading to new questions and possibilities. In AI, this could mean systems that not only solve predefined problems but also actively seek out new challenges and create innovative solutions, potentially surpassing human capabilities in various domains.
Provided that the real, significant challenges of AI safety and existential risk can be solved, there are tremendous gains to be had by creating more powerful AI. One of the most exciting prospects is for AI to automatically make groundbreaking discoveries that significantly enhance human life. Imagine AI systems that continuously innovate across various domains such as algorithm design, materials science, or protein folding. We are already beginning to see AI applied in these areas, but the future holds even more promise. These systems could potentially uncover new neural architecture designs or develop advanced sustainable materials — all at a pace and scale beyond human capacity alone. The key lies in developing AI that doesn't just process existing information but generates truly novel ideas and approaches, pushing the boundaries of human knowledge and capabilities.
The AI community has invested significant effort in studying open-endedness. If you're interested in diving deeper into this topic, check out this awesome list on open-endedness with many great contributions from the community. It's a great starting point for anyone looking to explore this exciting field further!
How was AiWallz inspired by open-endedness?
While my work has been influenced by numerous research endeavors, I'll focus on the main research project that inspired AiWallz. Recently, I collaborated on a research paper titled OMNI-EPIC: Open-endedness via Models of human Notions of Interestingness with Environments Programmed in Code. Our project focused on automatically generating learnable and interesting environments for AI agents, with the vision of enabling AI agent training in any learning environment. To progress towards this ambitious goal, we developed OMNI-EPIC, a system that leverages foundation models to autonomously generate code specifying tasks that are both learnable and interesting for AI agents.
This research project provided valuable insights for AiWallz, a project aimed at offering users a seamless experience in getting wallpapers that are both interesting and tailored to their preferences. We designed a simple user interface where users swipe left to skip a wallpaper or right to download it. Our algorithm learns from these interactions to generate new wallpapers that are more likely to align with the user's tastes. This approach draws inspiration from the concept of generating learnable and interesting tasks for AI agents, but applied in the context of generating the next interesting wallpaper that the user might enjoy.
While working on AiWallz, I encountered new questions and challenges. One key issue is accurately measuring the effectiveness of our algorithm in suggesting wallpapers that users prefer. Unlike AI agents, whose learnability is relatively easy to assess through performance on predefined tasks, evaluating user experience and preferences is more subjective. Other questions include: How does this approach compare with existing recommendation systems? What fidelity of user preference should/can we predict (e.g., the wallpaper's style and color scheme, or even down to the placement of objects in the wallpaper)? Finally, open-endedness has been extensively researched in AI systems, but how can we apply these concepts to user-facing applications like AiWallz?
There are many exciting directions that user-facing applications might benefit from by incorporating concepts of open-endedness, and we're excited to explore these possibilities further. If you're interested in learning more about AiWallz or have ideas on how to improve it, feel free to reach out to us!
written by Jenny Zhang