Imagine teaching an AI agent to navigate the web, not by tedious coding, but by simply feeding it online tutorials. That’s the promise of Synatra, a novel approach that transforms indirect knowledge, like how-to articles, into direct demonstrations for digital agents. Traditionally, training AI agents for complex web tasks requires massive datasets of specific actions paired with observations – data that’s expensive and difficult to acquire. Synatra bypasses this bottleneck by leveraging the wealth of readily available tutorials designed for humans. It breaks down a tutorial's steps, connects them to concrete actions within a web environment, and generates synthetic demonstrations at scale. Researchers used 100,000 such synthetic demonstrations to fine-tune a relatively small language model, CodeLlama-7b, and the results are impressive. The resulting agent, Synatra-CodeLlama, outperforms similar-sized models and even surpasses larger models like GPT-3.5 on certain web tasks. This success stems from Synatra's ability to teach the model the subtle art of web navigation. It learns to identify crucial details on web pages, like hidden links and buttons, and develops a clearer understanding of the logical flow of tasks. Remarkably, the cost of generating these synthetic demonstrations is only about 3% of collecting human demonstrations. This breakthrough suggests that with a powerful LLM, any ordinary online tutorial can secretly become a training ground for AI agents. Synatra's potential extends beyond simple web tasks, opening doors to training agents on a wider range of digital skills. This approach, though still with limitations such as dependence on source quality and computational cost, paves the way for more cost-effective, accessible, and efficient AI agent training, promising a future where AI can learn directly from the vast library of human knowledge.