Universe is a new AI training center that is supposed to teach computers to think more like humans.
Open AI, the nonprofit backed by Elon Musk and Peter Thiel to promote artificial intelligence
With Universe, developers can train artificial intelligence applications with games, websites, web browsers and other apps. The idea here is that the more an AI system practices using interfaces designed for human users, the more human-like AI can become.
But since Universe is open for anyone to use, that leaves the door open to developers who may utilize Universe to train AI in a way that would beget harm — precisely what Musk’s nonprofit aims to prevent. Opening developer tools is common in artificial intelligence programming, so the new initiative falls within standard practice.
Musk’s nonprofit is committed to open sourcing its tools and research as a way of hedging against the possibility of centralized, monopoly power over how artificial intelligence advances. Google opened its DeepMind artificial intelligence training tools on Monday, too. And earlier this year, Open AI released another open platform for training algorithms in complex environments called Gym.
This openness, however, potentially runs counter to Open AI’s purpose: To prevent Skynet from becoming self-aware.
“As our algorithms grow more sophisticated and our environments grow, we will be carefully thinking how to ensure people train AIs to ensure they have a good understanding of ethics, responsibility and culpability,” a spokesperson from Open AI told Recode. The team explored some of these issues in a recent paper and says they’re already working to build their own safe and secure systems.
Still, that doesn’t answer the question of how they plan to prevent developers from using Open AI’s free tools to build potentially unsafe and unethical artificial intelligence programs.
Universe aims to make AI better by becoming more adept at “general intelligence,” a concept within the AI community in which an AI learns a broader array of tasks, rather than being designed for one specific purpose.
Take Google’s AlphaGo, for example, the deep learning program that taught itself the ancient strategy game Go and defeated the best human player in the world earlier this year. AlphaGo’s win was considered a huge milestone in the development of artificial intelligence, but it wasn’t a demonstration of “general intelligence,” according to the team at Open AI.
Since the purpose of Universe is to edge AI closer to human-level intelligence, it uses virtual training environments that simulate how humans use computers by navigating the training exercises with mouse clicks and keyboard strokes.
Universe’s style of training might amount to better and smarter artificially intelligent programs, but that doesn’t mean it will necessarily amount to more human-like AI.
Open AI partnered with large game makers, like Microsoft and Valve, to provide about a thousand different video games for developers to train with. There are also environments for AI programs to practice using web browsers or spreadsheets, design with CAD or use a photo editing program.