Artificial intelligence may be “book smart,” but experts say it needs to become “street smart” to move beyond today’s limitations.
According to a recent Wall Street Journal report, the key to this evolution lies in developing AI world models — virtual environments that allow machines to learn and act more like humans and animals.
How Do AI World Models Work?
AI world models enable artificial intelligence to build an abstract understanding of its surroundings, plan actions and predict outcomes.
Instead of training only on static data such as text, images and videos, developers can use high-fidelity simulations to teach AIs how the world works — similar to learning to drive with “Gran Turismo” or to fly using “Microsoft Flight Simulator.”
Researchers believe this approach is crucial for advancing toward artificial general intelligence (AGI).
Leading figures, including Stanford professor Fei-Fei Li, have already raised major funding — Li secured $230 million for her startup World Labs — to accelerate work on these models.
Tech Giants Race
Tech giants are also racing ahead.
At Google DeepMind, scientists created Genie 3, a system capable of generating photo-realistic, open-world landscapes from a simple text prompt.
The project provides a virtual space where an “infant” AI can play, make mistakes and learn through reinforcement learning, just as humans do in real life.
Such environments could be used to train robots, self-driving cars and other “embodied” AIs to navigate complex spaces, interact with people and avoid obstacles before being deployed in the real world.
AI Much Better at Tasks Where it Currently Struggles
While it remains unclear whether AI world models will ultimately lead to the superintelligence some corporate leaders envision, experts agree on their near-term potential: making AI much better at tasks where it currently struggles, especially spatial reasoning and decision-making.
