AI Just Got Better at Real Tasks by Learning from Code Libraries

Imagine asking your AI assistant to set up a holiday trip: book flights, arrange a car, and reserve a hotel. Right now, it might struggle because it only knows what it learned from text and conversations. But researchers are teaching AI in a smarter way: by reading software code, which contains exact, tested instructions for how to complete actions. Their new system, called Code2Skill, looks at real programs on GitHub and breaks them into small, reusable "skills" — like how to check prices, compare options, or confirm a booking.

It does this at an enormous scale. The team scanned 19,769 popular, actively maintained code projects and created over one million verified skill records. Each one is checked to make sure it actually works. Because code is concrete and executable, these skills are far more reliable than skills learned from random web pages or past conversations with other systems. It's the difference between reading a recipe and actually watching someone cook.

When they tested AI models equipped with these skills across 72 different scenarios, the models improved by an average of 11.7% — and beat their previous best performance in 57 of those cases. Even more interesting, skills generated from AI-written code passed practical tests 93.5% of the time, almost identical to the 93% pass rate from human-written code. That suggests AI can keep teaching itself new skills as it creates more software.

So why should you care? AI assistants are moving from giving answers to taking action. That means asking your phone to sort your files, negotiate a refund, or manage your calendar becomes much more reliable. The days of AI saying "I'll explain how" may soon give way to "I've done it for you."