akshay Balsubramani

A central challenge of AI-driven biotech is that in drug discovery a fully validated medicine is worth a million times more than a million early-stage ideas that only have little validation. This is unfortunately at odds with the AI domain where we almost always attempt to "solve" a general class of problems - which even in the optimal case provides us at most with a little bit of validation for a large space of ideas, with most questions remaining unanswered and relegated to further downstream steps. Because of the large number of steps involved in drug discovery, solving even a general class of problems within just one of these steps does not massively impact the overall enterprise. This dichotomy is at the core of the challenges that first-generation data-generating AI-driven biotechs face - platforms produce a million low-validation ideas that then further clog up the validation bottleneck. Day to day, swimming in an ever-expanding lake of early-stage ideas may feel like progress, but the impact, measured in new medicines available for patients, is limited. Moreover, this approach to solve the general before the specific neglects the time factor - if you set out to solve a general class of problems before embarking on any specific idea, you will be overtaken by those that elect to instead address the valuable but specific problems in that space immediately. By the time your general capability is ready for action, you may find that the best ideas have already been pursued by others that did wait to first solve the entire class of problems. In short, we perhaps were too busy building platforms that we forgot to make medicines. What could we do differently to help address this central challenge? For one, there may be potential to focus more on the capabilities that enable us to produce fully validated or nearly fully validated ideas (depicted as a full glass of water) more quickly, that produce ideas of greater quality when fully validated, or perhaps even capabilities that open up new avenues that could not have been pursued otherwise. Secondly, it may help to mindfully balance general capability building with actually developing a medicine *today* to ensure we remain on track for realized and not only envisioned impact. The future of AI-driven drug discovery is bright but needs to be grounded in actual innovative medicines available for patients.