Thread by @JMateosGarcia

"Superhuman science"
This paper explores the potential impact of AI on scientific fields where innovation is about searching big spaces of possibility (e.g. new drugs, materials, architectures to build better AI systems...).
brookings.edu/wp-content/upl…
#econai ht @mattsclancy

1. AI systems trained on historical data can help predict which opportunities have more potential, improving prioritisation and expected returns from innovation.
But downstream blockers (e.g. around testing of the ideas that are developed) could constrain these benefits.

2. The paper gets into the weeds of this by modelling innovation as a multi-stage process of idea generation and testing where tests results at different stages are not wholly reliable. It would be cool to map and operationalise this idea in specific fields.

3. They also identify key resources needed for this to work: first of all, we need data of innovation successes and *failures* to train the AI system. Publication bias could mess this up. We also need interdisciplinary combinations of subject + ML / AI skills.

4. They also highlight a market failure: AI systems might be less useful for out-of-sample (radical) innovation in unexplored areas of science where there are less data.
This could create a bias towards incrementalism (consistent with results in Bianchini et al, 2022).

5. Government can help address this market failure by funding risky research that explores new areas of science and creates data to train improved AI systems that can be used in follow-on work.
/ends