Breakthrough in AI Technology Indicates Potential for Realistic Machine Understanding

=========================================================

In a groundbreaking development, researchers from MIT have proposed a new approach to improve the reasoning abilities of large language models. Known as the "libraries of abstraction," this project aims to provide human-like reasoning for functions such as computer programming, task planning, and robotic tasks.

The team behind this initiative includes Aviral Kumar, Chelsea Finn, Sergey Levine, and Gabe Grand, an MIT PhD student who serves as a lead author on one of the research papers. Their findings were presented at the prestigious NeurIPS 2023 conference and more recently at the International Conference on Learning Representations in Vienna.

The focus of this study was on household tasks and command-based video games. The researchers demonstrated the complex background reasoning of the human mind, a challenging aspect to replicate in AI, through three libraries: LILO, Ada, and LGA.

LILO, which uses a neurosymbolic method and the Stitch algorithm, identifies abstractions, allowing for the application of commonsense knowledge with sophistication. Ada, on the other hand, showcases the beauty and power of the human mind, a concept that is difficult to replicate in AI.

The libraries of abstraction create more straightforward abstractions for language models and assign natural language names and documentation to each one, leading to more interpretable code for programmers. This is significant because language models prefer to work with functions that are named in natural language.

When tested on Boston Dynamics' robot Spot, the language models were able to create a plan of action in an "unstructured environment" and complete tasks beyond simple image recognition. Moreover, implementing these techniques with GPT-4 improved the accuracy of AI actions in tasks such as making breakfast or playing Minecraft.

The press statement from MIT refers to the libraries of abstraction as a significant step towards imparting something akin to human reasoning on AI. They also emphasise that these libraries provide a rich source of context for language models.

Darren Orf, the author of this article, resides in Portland and is known for writing and editing about sci-fi and the workings of our world. His previous works can be found at Gizmodo and Paste.

This kind of task navigation could have real-world implications for driverless cars or other autonomous technologies. As we continue to advance in AI development, the libraries of abstraction represent a significant advancement in this exciting field.