Imagine a world where AI could design life-saving drugs, optimizing molecular structures for maximum effectiveness against diseases. That future might be closer than you think. Researchers are exploring the use of Large Language Models (LLMs), typically used for text analysis, in the realm of molecular optimization. This involves navigating the vast and complex chemical space to find the perfect molecule for a given task, such as a drug that targets a specific disease protein. Traditionally, this has been a laborious process of trial and error in the lab. Enter LICO, a new method that empowers LLMs to tackle this challenge. LICO works by adding specialized components to the LLM, enabling it to understand and ‘reason' about molecules and their properties. Instead of feeding the LLM lengthy textual descriptions, LICO represents molecules as compact vectors, allowing the model to process more information efficiently. This allows the LLM to learn from past experiments and make informed predictions about how changes to a molecule's structure will affect its properties. To train LICO, researchers expose it to a mix of ‘intrinsic’ and ‘synthetic’ functions. Intrinsic functions are readily calculable properties of the molecule like its weight or number of rings. Synthetic functions, generated by a mathematical model called a Gaussian Process, introduce additional complexity and diversity, mimicking the kind of functions scientists might want to optimize in the real world. This combined approach enables LICO to generalize its knowledge and optimize for a wider range of molecular objectives. Tested against existing molecular optimization tools, LICO comes out on top. On a benchmark called Practical Molecular Optimization (PMO), which features a variety of objective functions, LICO consistently shows it can achieve better results with fewer experimental trials, crucial for speeding up and lowering the cost of drug discovery. While the results are promising, there are still hurdles to overcome. One is the assumption that researchers have access to a set of relevant intrinsic functions. For molecules, this works well, but in other scientific domains, such functions might not be readily available. Future research will focus on generating high-quality synthetic data that encodes essential domain knowledge, bridging the gap where intrinsic functions are lacking. Scaling up LLM size also seems to improve performance, so leveraging even larger models holds potential for further gains. LICO provides a glimpse into the exciting potential of LLMs in science, moving beyond text to address complex optimization problems with real-world implications. This could accelerate the development of new drugs and materials, ushering in a new era of AI-driven scientific discovery.