AI startup DeepMind, acquired by Google in 2014 and most famous for its crushing prowess in games like Chess and Go, is setting its sights on real scientific problems, but its successes and failures point to how much data is really required for modern AI to accomplish anything meaningful. Their current project is to predict protein folding (this is important for a lot of things, like targeting cancer cells, or proteins that cause Alzheimer’s, for example). There’s been an explosion of genomic data since 2006 (expected to exceed 40 exabytes of storage by 2025). This kind of data is central to “machine learning”. With what’s available now, the DeepMind team was able to compete in the 2018 CASP (Critical Assessment of Techniques for Protein Structure Prediction), where they correctly predicted 25 of 43 proteins modeled from scratch. Fifty-eight percent doesn’t necessarily sound great, but they were more than 8 times more successful than the next closet team. This both points to the power of AI, and how much farther it has to go.