L2MAC: Large Language Model Automatic Computer

Pioneering the first practical LLM-based general-purpose stored-program automatic computer (von Neumann architecture) framework in an LLM-based multi-agent system, for solving complex tasks through generating extensive and consistent outputs, unbounded by the LLMs fixed context window constraint.

LLM-Automatic Computer (L2MAC) framework excels at solving large complex tasks, such as being state-of-the-art for generating large codebases, or it can even write entire books, all of which bypass the traditional constraints of the LLMs fixed context window constraint.

• LLM-Automatic Computer can take a one line input prompt and generate an extensive and large output, for example generating an entire complex codebase.
  • Internally, the LLM-Automatic Computer uses a prompt-program which is a series of prompts, each providing a instruction step to execute. Unless explicitly given, the prompt-program is self generated (bootstrapped) and executed. Specifically each instruction step of the prompt-program is loaded into a new LLM agent to execute, whose context is managed by a control unit and is provided with tools so that it can read and write to a persistent memory, here a file store, which contains the final and intermediate outputs. This enables it to automatically execute general-purpose prompt programs to solve complex tasks, that require extensive cohesive outputs, where the output is unbounded and not constrained by the LLMs underlying context window constraint.

LLM-Automatic Computer (L2MAC) instantiation for coding a large complex codebase for an entire application based on a single user prompt. Here we provide L2MAC with additional tools to check for any syntax errors within the code and run any unit tests if they exist, and call this instantiation Code-L2MAC.

L2MAC's Abilities

LLM-Automatic Computer (L2MAC) is an LLM-agent framework created within the University of Cambridge van der Schaar research lab, emanating from the peer-reviewed published paper in ICLR 2024. You can use this multi-agent framework to solve your complex task, and create your own full code application or large text outputs, such as writing books or reports. For more details, you can refer to CodeBase Generator and Book Generator under Use Cases. Let us start with a complete example.

Examples (fully generated by GPT-4)

For example, if you type l2mac "Create a beautiful, playable and simple snake game with pygame. Make the snake and food be aligned to the same 10-pixel grid.", you would get a complete codebase for a fully playable game. See the generated codebase at CodeBase Generator.

This example costs around $0.16 for the complete codebase repository.