You need to sign in or sign up before continuing.
Take a photo of a barcode or cover
evan_reads 's review for:
AI Snake Oil: What Artificial Intelligence Can Do, What It Can't, and How to Tell the Difference
by Sayash Kapoor, Arvind Narayanan
Many interesting and very relevant points. New things that stuck out to me:
- reproducibility crisis in ML extends from academia to industry (e.g. Epic sepsis model, startups gaming their eval metrics for funding, VCs wanting to increase hype) but with even less ability to investigate due to non-transparency
- Top-N accuracy "hacking" to achieve 90% accuracy
- Exploitative labor required to train generative models (MTurks)
- case for partial lotteries over current opaque and arbitrary selection systems
- Content moderation follows cycle of making/revising policies, enforcement, appeal, and evaluation. AI can only help with the enforcement part
- reproducibility crisis in ML extends from academia to industry (e.g. Epic sepsis model, startups gaming their eval metrics for funding, VCs wanting to increase hype) but with even less ability to investigate due to non-transparency
- Top-N accuracy "hacking" to achieve 90% accuracy
- Exploitative labor required to train generative models (MTurks)
- case for partial lotteries over current opaque and arbitrary selection systems
- Content moderation follows cycle of making/revising policies, enforcement, appeal, and evaluation. AI can only help with the enforcement part