AI Roundup 11/24 -> 12/01/2022

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

1. OpenAI releases ChatGPT - a new AI system optimized for dialogue and reducing disinformation & hallucinations by fact-checking, asking clarifying questions, and admitting mistakes (OpenAI)

2. Researchers from Meta AI and the Hebrew University of Jerusalem present SpaText - a new text-to-image generation method to generate new images from a 'global' prompt defining the scene along with a segmented part of an image and another 'local' prompt to replace the segmented part (Paper's website)

3. OpenAI introduces a new GPT-3 model text-davinci-003, which generates complex and long-form content like poetry and songs rhyming using fewer instructions compared to previous models (Ars Technica)

4. Google researchers present two methods for improving the performance of existing language models by a large margin without using massive computational resources (Google AI)

5. CarperAI announces the release of OpenELM (Evolution Through Large Models) - an open-source library combining LLMs with evolutionary algorithms for code synthesis (CarperAI)

6. DeepMind introduces a new framework that uses Reinforcement Learning to build multimodal interactive agents capable of following instructions and performing actions in open-ended conditions such as in video games (DeepMind)

7. Researchers from ByteDance introduce Corgi, a novel general diffusion model implemented on text-to-image generations, achieving SOTA results, outperforming DALL-E 2 in both image quality and efficiency (Paper)

8. Together, an open-source AI community, releases GPT-JT - a fork of GPT-J outperforming other 100B+ parameter models at classification (Analytics India Magazine)

9. Researchers from Microsoft and the City University of Hong Kong introduce a new framework to improve the reasoning and common sense abilities of recent vision-language models (Paper)

10. Hugging Face collaborates with arXiv Labs to help discover AI-based research papers, linking to demos created by the community (Hugging Face)

11. OpenAI introduces the best Minecraft-playing bot to date by training using ~70K hours of human gameplay, leveraging a technique called Imitation Learning (MIT Tech Review)

12. Researchers from Weizmann Institute of Science present a new framework that leverages diffusion features to image-to-image translation, capable of injecting extracted features from the original image into the target one (Paper's website)

Last week’s AI Tidbits