... Toner-Rodgers told a story of a thousand material science researchers, at an unnamed company, who used a machine learning system to generate possible new materials. With the AI, they found 44% more new materials, patent filings went up 39% and new product prototypes went up 17%. Incredible! Though he did say the scientists felt alienated from their work.
The paper exploded. Economists loved it! Toner-Rodgers submitted the paper to the Quarterly Journal of Economics.
More importantly, the paper told the AI promoters what they wanted to hear — the bosses don’t care about the disgruntled workers, but they do really want more output...
Robert Palgrave, a professor of inorganic chemistry and materials science at UCL, goes through a pile of things about the paper that struck him as not quite right, both in December 2024 and just recently:
The original paper was too good to be true in many ways. How could a 2nd year PhD student get pervasive access to extremely sensitive data from what must have been a multi billion dollar company?
How could such a study have been set up years before this student started his work PhD, in just the right way to deliver results to him?
What company really has 1000 scientists all trying to discover new materials all day every day? It didn’t really make sense as a concept.
But there were technical points too. AI, using atomistic methods like DFT, cannot predict most of the types of materials that were supposedly studied. It can only really work for simple crystalline materials. Glasses, biomaterials? No chance...