The day started at 6am PST with the Black in AI workshop, and the fascinating posters at 9pm PST included GPT-3 and Big Bird 😪☕
One of the prominent members in the community, who won the Turing award for his contributions herein, saw a super-resolution algorithm turn a pixelated presidential portrait of Obama into a generic white guy, and said it was biased data, and the first invited talk took an hour to refute that.
- AI as a community needs to be software engineers and language nerds and ethnographers as well
There’s an unfathomably wide array of applications (even avian note segmentation!)
The ergonomics of using quick convenient at-hand devices over expensive “best practices” is interesting
- Kwame (by Jojo Boateng) is NLP for question answering for SuaCode (coding mostly by phone)
- Denoising ultrasounds that can plug into phones by Muyinatu Bell
- Apima(?): efficient malaria dataset generation (via images from an iPhone 6s+ (5 y/o in 2020!) of stained blood in a microscope)
- Deploying webapps straight from Jupyter notebooks
Major NLP advances!
- huge prompt contexts, 3K tokens in Big Bird
- GPT-3, with those 175B parameters
- Lesan translates English to Amharic better than Google Translate
- transfer learning Zulu via Xhosa and Shona
- long models for molecular representation learning
AR Clip Drop started as a weekend prototype, and Cyril Diagne had projects for the weekend after lol but those got delayed after trying to productionize the solo-type-of-prototype