Training AI models often requires tons of labeled data, a process that can be eye-wateringly expensive. Imagine paying thousands of dollars just to label a modest dataset – that's the reality many AI developers face when using large language models (LLMs) for annotation. But what if there was a cheaper, more efficient way? Researchers have unveiled a groundbreaking new system called "Alchemist" that could revolutionize data labeling. Instead of asking LLMs to label every single data point, Alchemist prompts them to write small programs that can do the labeling automatically. These programs can then be run locally, eliminating the need for costly API calls to LLMs. The result? A staggering 500x reduction in labeling costs, along with a 13% average improvement in accuracy across various tasks. This innovative approach addresses several key challenges in LLM-based annotation. High costs, lack of flexibility to adapt labeling rules, and the 'black box' nature of LLMs make traditional annotation methods difficult to manage. Alchemist tackles these issues head-on. Its generated programs are transparent, allowing for easy auditing and modification. They can be reused and extended, providing a dynamic labeling solution. Even better, Alchemist can handle diverse data types like text and images, boosting its versatility. The system works by prompting LLMs with a task description and instructions on the desired output format. For images, Alchemist cleverly extracts high-level concepts, uses a local model to convert images into feature vectors, and then prompts the LLM to generate a program based on those features. While Alchemist's performance is tied to the LLM's capabilities, the system offers a powerful new paradigm for affordable and efficient AI data labeling, opening doors for wider adoption of AI across various industries.