Researchers at DeepMind, Princeton University, and Stanford University have recently proposed an innovative framework called LATM (LLMs as Tool Makers) that transforms large language models like GPT-4 into tool makers.
The innovation of LATM is that it introduces the concept of tool making and use to language models, enabling them to autonomously create tools for different tasks. This framework not only increases the flexibility and adaptability of models but also provides a more efficient and cost-effective solution for generative AI.
The framework has two main components:
- Tool making: As a tool maker, LLM specializes in designing software tools for specific tasks, which are implemented as Python functions.
- Tool usage: Another LLM, as a tool user, can invoke these software tools to process new requests.
This design allows LATM to assign tasks to the most appropriate LLM, assigning tool-making processes that require high computational power to powerful, resource-intensive models, such as GPT-4, and assigning relatively simple tool-using processes to lightweight, cost-effective models, such as GPT-3.5 Turbo. This not only enhances problem-solving capabilities but also significantly reduces the average computational cost of processing a range of tasks to maximize the efficiency of the framework.