new paper: can AI forecast scientific progress, or is it just good at scientific hindsight?
the authors introduce CUSP: 4,760 real scientific events across biology, chemistry, physics, medicine and AI, with historical knowledge cutoffs so models only see what was available before the discovery
tasks include:
- will this advance happen?
- what mechanism explains it?
- can you propose the solution?
- when will it appear?
Great work by @sean_n_wu @philiptorr @james_y_zou
@lupantech @turingmusician @_yutaroyamada