AI Roundup 03/09 -> 03/16/2023

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

1. OpenAI releases GPT-4, a multimodal AI outperforming its predecessor ChatGPT, scoring higher across multiple simulated exams including SATs and BAR Exams (OpenAI)

   

2. Google takes on ChatGPT by announcing the PaLM API and MakerSuit, allowing developers to build and prototype generative AI apps using its SOTA LLM (Google Blog)

3. Google announces MedPaLM2, a medical LLM, that achieved a score of 85% on a medical exam (USMLE MedQA), surpassing GPT-4 and is on par with an expert doctors (Google Blog)

4. Anthropic introduces Claude, its LLM AI assistant, and provides access through a chat interface and API (Anthropic Blog)

5. Together releases OpenChatKit, an open-source 20B parameter ChatGPT-like model that can reason and converse (TechCrunch)

6. OpenAI open sources Evals, a framework that leverages customer data to train and help evaluate the performance of its AI models (Github repo)

7. Midjourney releases V5 for its text-to-image generation model having better image quality, more diverse outputs, wider stylistic range, and better image prompting (Twitter)

   

8. Researchers propose Tag2Text - a vision language framework that improves the learning of visual-linguistic features by introducing image tagging (Paper's Github repo)

8. Researchers from Microsoft introduce MathPrompter - a novel approach that improves LLMs' performance on mathematical reasoning achieving state-of-the-art results on the MultiArith dataset (Paper)

9. Researchers from Seoul National University introduce a novel text-to-video editing model (Paper's Github repo)

10. Researchers from the Hebrew University of Jerusalem introduce WHOOPS! - a dataset challenges AI models to identify and explain why unconventional images are unusual (Paper's Github repo)

11. Adobe researchers present 3D Cinematography, transforming a single 2D image into a 3D visual animation with camera motion (Paper's Github repo)

12. Berkeley and Microsoft showcase V-MoE - a scalable and efficient image recognition model matching the performance of state-of-the-art dense networks with half the compute at inference time (Paper)

13. A group of researchers present UniDiffuser a diffusion-based model capable of generating realistic samples in various tasks, including image, text, text-to-image, image-to-text, and image-text pair generation (Github repo)

14. Meta released Casual Conversations v2, a video dataset from various countries, to test the fairness and reliability of computer vision and speech recognition models (Meta AI)

Last week’s AI Tidbits