,
12 tweets,
6 min read
Read on Twitter
[Highlights of day 2 of the Economics of AI @nberpubs workshop] Sonny Tambe presented a cool paper using LinkedIn data to recover firm investments in intangible AI-related assets, finding it concentrated in 'superstar' firms (FAO @stianwestlake)
papers.ssrn.com/sol3/papers.cf…
papers.ssrn.com/sol3/papers.cf…
3. I presented our paper mapping Deep Learning research & drivers of regional advantage in AI research w/ arXiv data.
Paper: arxiv.org/abs/1808.06355
Code: github.com/nestauk/arxiv_…
(@FYI @jackclarkSF you asked for this)
Slides: slideshare.net/JuanMateosGarc…
Paper: arxiv.org/abs/1808.06355
Code: github.com/nestauk/arxiv_…
(@FYI @jackclarkSF you asked for this)
Slides: slideshare.net/JuanMateosGarc…
4. I already tweeted about this paper modelling the impact of AI in scientific search.
4b. This slide shows the combinatorial plurality of methodological combinations in the discovery process. Maps & experiments are complements, not substitutes (FAO @IGLglobal). The idea of modelling all this using reinforcement learning sounds very cool. 
5. I have said before that we need an AI Growth Lab that uses RCTs to systematically & openly test the effectiveness of different org designs to implement AI. This Glaeser et al paper about health inspection recommenders is a great example of how this could work.
5b. Interesting finding: inspectors tend to not-comply with algorithmic recommendations even though they perform much better than alternative models. There are many potential reasons for this, illustrating the complexity of AI deployment 'in the wild'.
6. Very cool macro paper by Benzell et al of yet another way to reach AI-mediated dystopia, this time through the impoverishment of younger generations leading to a collapse in savings & investments. Based on an Isaac Asimov story (yay!)
conference.nber.org/conf_papers/f1…
conference.nber.org/conf_papers/f1…
7. When discussing this session on AI & jobs, Jason Furman called for:
+ empirical work & international comparisons complementing theoretical models
+ attention to understanding what worked in the recent past complementing the obsession with the future of work.
+ empirical work & international comparisons complementing theoretical models
+ attention to understanding what worked in the recent past complementing the obsession with the future of work.
8. Does the use of algorithms in mortgage decision-making replace human prejudices with correlational ones? Very powerful, beautiful work by Adair Morse & colleagues, combining econ & law & using millions of records & ML-powered econometrics.
conference.nber.org/conf_papers/f1…
conference.nber.org/conf_papers/f1…
8b. In a nutshell, fintechs using AI to make mortgage decisions remove human prejudice but introduce discrimination through correlates of protected characteristics. The net effect is positive (less discrimination) but still worrying (is it the tip of the iceberg?)
9. And that's it! There were a couple more papers but jet-lag had taken over by then. Congrats to the organisers for a inspiring conference. Looking forward to see where the community goes next.
