AI Roundup 02/17 -> 02/23/2023

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

1. AWS publishes Multimodal Chain-of-Thought Reasoning, outperforming GPT-3.5 by 16% (75%->91%) and surpasses human performance on the ScienceQA benchmark with less than 1B params (Paper's Github repo)

2. Meta researchers propose LEVER, a state-of-the-art approach that improves language-to-code generation by learning to verify the generated programs with their execution results (Paper)

3. Researchers from Meta AI and the University of British Columbia present MINOTAUR - a unified multi-task model for query-based video understanding (Paper)

4. Microsoft researchers introduce a novel self-supervised diffusion model, which only trains on single images but can generate high-quality 3D videos (Paper)

5. UC San Diego and Nvidia propose ACE-VC - a zero-shot voice conversion model that can mimic a target speaker (Paper's website)

6. Stanford researchers present Hyena - a new method achieving state-of-the-art performance on language tasks while reducing training costs by 20% and improving inference time by up to 100x (Paper)

7. Researchers from the Weizmann Institute of Science introduce MultiDiffusion - a unified framework that enables versatile and controllable image generation, using a pre-trained text-to-image diffusion model (Paper's Github repo)

   

8. Researchers from CMU introduce a model capable of generating high-resolution 3D photorealistic images from a 2D label (Paper's Github repo)

   

9. Alibaba researchers present Composer - a new generation paradigm that allows flexible control of the output image, such as spatial layout and palette (Paper)

10. Researchers from Oxford demonstrate a state-of-the-art model called RealFusion, capable of reconstructing a full 360° photographic model of an object from a single image (Paper's Github repo)

11. Roblox expects to launch AI tools for users to create and change the look of in-game materials with textual prompts (Wired)

12. The Supreme Court stated that tech companies building generative AI applications might be liable for their outputs (VentureBeat)

Last week’s AI Tidbits