Morgan Levine asks what business model would let AI learn general patterns in biology before there is a product

Morgan Levine asks what business model would let AI learn general patterns in biology before there is a product

On October 23, Morgan Levine shared a thought experiment in a post on X. Would early large language models have received funding if investors had required a customer, a narrowly defined application, and near-term revenue from the outset?

Levine writes that such requirements could have steered developers toward solutions with clear, immediate benefits.

In her account, developers began training language models to learn general patterns in data through self-supervised learning, in which the training signal comes from the data themselves. Self-supervised learning allowed the models to develop capabilities beyond predefined tasks. Those capabilities later created commercial value that would have been difficult to anticipate.

Levine suggests that a requirement for near-term revenue can determine which research topics are pursued before the work even begins.

She extends this reasoning to AI for biology. Projects with clear near-term applications may be selected for funding before researchers can establish what might come from developing general representations of biological systems.

Levine describes venture capital as a successful way to turn technological breakthroughs into companies. For basic research, she asks about an additional business model:

“Can we create commercial funding models that give basic research the time it needs while preserving the incentive to put its findings into practice?”

Levine closes her post by asking:

“What transformative ideas are we missing because they do not fit existing models?”