AI Roundup 11/10 -> 11/17/2022

Last week's most interesting AI news in <2 minutes

1. Meta AI open sources Galactica, a 120B parameters model trained on 48M scientific texts capable of conducting tasks such as solving math problems, reasoning, generating documents, etc. (Meta AI)

2. A group of researchers introduces Constructive Search - a new decoding method achieving human-level text generation performance (Hugging Face)

3. Deepmind researchers introduce TAP-VID, a new benchmark to track physical surfaces in videos, along with TAP-Net, an end-to-end point tracking model outperforming prior methods (Twitter)

4. GitHub researchers preview AI for Pull Requests - a set of functionalities to automatically review and resolve pull requests using language models (Project website)

5. Boston University and Google researchers introduce a model illustrating articles with visual images. The release includes NewsStories - a multimodal dataset with >31M articles, 22M images, and 1M videos (Paper's Github repo)

6. Researchers from Microsoft and TUDelft introduce an AI-based approach for generating synthetic expression-based face wrinkles (Marketpost)

7. Metaphor releases a new search engine powered by generative AI to surface URLs from a given prompt (Twitter)

8. Canva launches a text-to-image feature powered by Stable Diffusion to its >100M users to compete with Microsoft's Designer (The Verge)

Last week’s AI Tidbits