Accuracy in AI is one of the most widely cited concerns of the technology. But as Sam Altman from OpenAI says “the tools you’re using now are the worst ones you’re ever going to use.”
In other words, things are getting a lot better. However it’s going to definitely take real solid progress to win back some of the distrust that’s been established from some high profile cases where things have gone (very) awry!
Well, this may be a good step in the right direction…
Artificial intelligence (AI) planning is all about crafting a sequence of actions to achieve a specific goal, and it’s crucial in fields like robotics and logistics. Recently, researchers from Cornell University and IBM Research have introduced a new system called AutoToS, which aims to automate the generation of sound and complete search components in AI planning. This development is particularly significant because it reduces the need for human intervention, which has been a major bottleneck in creating scalable AI planning systems.
How AutoToS Works
AutoToS is designed to generate search components that are both sound and complete without human oversight. It uses large language models (LLMs) to extract successor functions and goal tests, then automatically tests these components using unit tests. If any part of the generated code doesn’t meet the criteria for soundness or completeness, AutoToS provides detailed feedback to the LLM for revisions. This iterative process continues until the components are fully validated. The system employs both Breadth-First Search and Depth-First Search with additional checks to ensure the search process is sound and complete.
Benefits
- Increased Scalability: By automating the feedback process, AutoToS significantly reduces the need for human intervention, making it easier to scale AI planning systems.
- High Accuracy: In tests, AutoToS achieved 100% accuracy in domains like BlocksWorld, PrOntoQA, Mini Crossword, the 24 Game, and Sokoban.
- Efficiency: The system required fewer feedback iterations to achieve perfect results, averaging just 2.6 calls to the LLM for the 24 Game domain and 2.8 calls for BlocksWorld.
Concerns
While AutoToS shows great promise, there are potential concerns. The system’s reliance on LLMs means it could be affected by the limitations of these models, such as biases in the training data. Additionally, the initial setup and configuration of AutoToS might require significant effort and expertise.
Possible Business Use Cases
- Automated Logistics Planning: Use AutoToS to optimize supply chain and logistics operations, reducing the need for human planners.
- Robotics Navigation: Implement AutoToS in autonomous robots to improve their navigation and task planning capabilities.
- Game Development: Leverage AutoToS to create more intelligent and adaptable AI opponents in video games.
As we move towards more automated and scalable AI planning systems, how can we ensure that the technology remains transparent and free from biases?
Image Credit: DALL-E
