AI Roundup 05/18-> 05/25/2023

A model called DragGAN to manipulate images, Meta's multilingual speech model supporting >1k languages, a ChatGPT-like AI for video editing, and Google's multimodal model to automate web navigation

1. Researchers from Google, MIT, and Max Planck Institute introduce DragGAN, a new model capable of manipulating the pose, shape, expression, and layout of diverse image categories (Paper)

2. Microsoft researchers present CoDi, a novel model capable of generating various combinations of output modalities from diverse inputs like video, image, and text (Paper's website)

3. Researchers from the University of Cambridge present PandaGPT, the first general-purpose multimodal model capable of following instructions across six modalities (Paper's website)

4. Meta open sources Massively Multilingual Speech (MMS), a model supporting speech-to-text and text-to-speech for >1,100 languages (Meta AI)

5. SambaNova and Together present BLOOMChat - an open-source multilingual chatbot with a commercial permissive license, achieving a win rate of 45.25% compared to GPT-4's 54.75% in a human preference study (Paper's website)

6. Google Brain presents WebGUM, a novel method for training agents to perform web navigation tasks like typing and clicking using a combination of language and visual models, surpassing previous methods (Paper)

7. Meta AI unveiled LIMA - a fine-tuned version of their 65B LLaMA achieving Bard level performance and having strong generalization capabilities for unseen tasks (Paper)

8. Researchers from Sun Yat-Sen University introduce Video-ControlNet, a text-to-video diffusion model capable of generating high-quality videos while allowing fine-grained control over the generated content (Paper's website)

9. Researchers from multiple renowned universities introduce RWKV - a novel neural network architecture combining the strengths of RNNs and Transformers while overcoming their limitations (Paper)

10. Researchers from the National Uni of Singapore and The Chinese University of Hong Kong introduce Mind-Video, a novel approach to reconstructing human vision from brain activities (Paper's website)

11. Microsoft researchers introduce Pengi, an Audio-Language model that generates text from a combination of audio and textual prompts (Paper)

12. Stanford researchers introduce AlpacaFarm, a simulator that enables fast and cost-effective research on methods of learning from human feedback, addressing challenges like high annotation cost and lack of evaluation (Stanford Research)

13. Meta introduces MegaByte, a new architecture that improves Transformer models in both cost and performance by replacing tokens with bytes (Analytics India Magazine)

14. Google presents a new framework to automate and streamline the code review process using ML (Google AI)

Announcements

1. OpenAI releases a free ChatGPT iPhone app in the U.S, supporting voice input (OpenAI)

2. Adobe announces Generative Fill, a new AI-powered image generator integrated into Photoshop, allowing users to add or delete elements using text prompts (The Verge)

3. Microsoft add native Bing support within ChatGPT, powering it with access to the web including citations and advanced search capabilities (Analytics India Magazine)

Thought-provoking

1. Bill Gates predicts that personalized AI Agents will alter users’ behaviors and will redefine Google Search and Amazon (CNBC)

2. Former Google CEO Eric Schmidt believes AI represents an "existential risk" that could cause harm to many people (WSJ)

Last week’s AI Tidbits

If you find AI Tidbits valuable, share it with a friend, and consider showing your support.