AI Roundup 01/12 -> 01/19/2023

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

Happy New Year everyone and welcome to the first issue of AI Tidbits for 2023. Judging by how it started and the multiple hints from AI celebrities—this year is going to be at least as wild as the previous one.

1. Deepmind introduces DreamerV3, the first multi-domain model that solves Minecraft's diamond challenge with no human data (Twitter)

2. Microsoft releases VALL-E, a zero-shot text-to-speech model capable of synthesizing high-quality speech from a 3-second recording (Project website)

3. AnthropicAI publishes a paper on LM-written evaluations where they uncover emerging behaviors such as inverse scaling (LMs get worse with size) (Anthropic)

   

4. Anthropic is also providing beta access to Claude - its ChatGPT-like model with a sneak peek comparing the two (Scale AI blog)

   

5. Deepmind considers releasing its ChatGPT rival Sparrow in private beta in 2023, stating it outperforms ChatGPT while adding citation capabilities (Yahoo!)

6. Researchers from Microsoft and Columbia University introduce GLIGEN - a novel text-to-image diffusion approach that enables generation with grounding inputs (Paper's website)

6. Researchers from Carnegie Mellon, Reality Labs, and Meta introduce HyperReel, a new method to render immersive high-quality videos with different viewing angels from given images (Paper's website)

7. Researchers from Deepmind present LASER-NV - a generative model for high-fidelity novel view synthesis of 3D scenes from arbitrary viewpoints (Paper's website)

8. Google reveals its vision for language, computer vision, multi-modal models, and generative AI for 2023 and beyond (Google AI)

9. AI21 Labs releases Wordtune Spices, a text-generating AI assistant that suggests improvements and backs up its suggestions by linking to citations (TechCrunch)

10. Researchers from Show Lab and the National University of Singapore introduce Tune-A-Video, a one-shot text-to-video generator (Analytics India Magazine)

11. Nvidia and Tel Aviv University unveil a new text-to-image method to alter existing images by embedding images as new words (Paper's website)

12. Meta partners with Shutterstock to accelerate the development of generative models using Shutterstock's collection of millions of images, videos, and music (Analytics India Magazine)