Funding Biology Before a Product Exists

Who Funds Biology Before a Product Exists

Morgan Levine asks what commercial model would give artificial intelligence time to learn general patterns in biology before its eventual product is known. On October 23, she proposed a thought experiment on the social platform X: would early language models have received funding if investors had required an identifiable customer, a narrow application, and near-term revenue from the outset?

In Levine's account, those requirements could have directed developers toward tasks with clear, immediate benefits. Developers trained language models to learn general patterns in data. The training signal came from the data themselves, an approach called self-supervised learning. Levine writes that this gave the models capabilities beyond predefined tasks. Those capabilities later acquired commercial value that would have been difficult to anticipate.

She extends this question to biology. If funding primarily goes to projects with near-term applications, researchers may lose the opportunity to discover what models could learn about general patterns in biological systems. Revenue requirements would then shape the choice of research topics before the work begins. You can find the author through her Eternal Search profile.

Levine also credits venture capital, investment in young companies with high risks, with successfully turning technological breakthroughs into companies. For basic research, she asks for an additional commercial model that gives the work enough time while preserving an incentive to put its findings into practice.

The news item offers no finished funding model. The alternative history of language model development remains a thought experiment. It explains Levine's concern: selecting projects for near-term revenue may leave research unfunded when its applications have yet to be discovered.

Open the related Eternal Search page

Sources
[1] x.com