Five Small AIs Beat One Big AI at Coding—Here's How

A team of five small AI models, each with a specific job, has outperformed a single large AI model on a set of difficult coding challenges. The method, called Memdoor, assigns roles like planner, idea generator, solver, and manager. Each AI works independently and communicates only through files, not chat. This division of labor allows the team to tackle complex problems more effectively. In tests on ten hard coding problems, the team solved 49 out of 50 attempts, with each attempt costing about 39 cents.

The key advantage is cost and efficiency. Instead of relying on one expensive, massive AI model, you can use several smaller, cheaper models that together achieve better results. This could make advanced coding assistance more accessible to small businesses and individuals. For example, a freelance developer could use this approach to get high-quality coding help without paying for a premium AI service. The method also allows for flexibility, as you can swap in different AI models as needed.

However, there are limitations. The tests were run on only ten problems, and the AI may have seen these problems during training, which could inflate the results. Also, the comparison with the large model was not run under identical conditions. Despite these caveats, the approach shows promise for making AI-powered coding more affordable and effective. The method is open and reproducible, meaning others can build on it.